Instructions for submitting a technical report or thesis.
Title: Improving Event Extraction: Casting a Wider Net
Candidate: Cao, Kai
Advisor(s): Grishman, Ralph
Abstract:
Information extraction is the task of automatically extracting structured information from unstructured and/or semi-structured machine-readable documents. One facet of information extraction is event extraction (EE): identifying instances of selected types of events appearing in natural language text. For each instance, EE should identify the type of the event, the event trigger (the word or phrase which evokes the event), the participants in the event, and (where possible) the time and place of the event.
One EE task was defined and intensively studied as part of the ACE (Automatic Content Extraction) research program. The 2005 ACE EE task involved 8 types and 33 subtypes of events. For instance, given the sentence "She was killed by an automobile yesterday.", an EE system should be able to recognize the word "killed" as a trigger for an event of subtype DIE, and discover "an automobile" and "yesterday" as the Agent and Time arguments. This task is quite challenging, as the same event might appear in the form of various trigger expressions and an expression might represent different types of events in different contexts.
To support the development and evaluation of ACE EE systems, the Linguistic Data Consortium annotated a text corpus (consisting primarily of news articles) with information on the events mentioned. This corpus was widely used to train ACE EE systems. However, the event instances in the ACE corpus are not evenly distributed, and so some frequent expressions involving ACE events do not appear in the training data, adversely affecting performance.
The thesis presents several strategies for improving the performance of EE. We first demonstrate the effectiveness of two types of linguistic analysis -- dependency regularization and Abstract Meaning Representation -- in boosting EE performance. Next we show the benefit of an active learning strategy in which a person is asked to judge a limited number of phrases which may be event triggers. Finally we report the impact of combining our baseline system with event patterns from a system developed for a different EE task (the TABARI program). This step contains expert-level patterns generated by other research groups. Because the information received is complicated and quite different from the original corpus (ACE), the integration of this information requires more complex processing.
Title: On the Design of Small Coarse Spaces for Domain Decomposition Algorithms
Author(s): Dohrmann, Clark; Widlund, Olof
Abstract:
Methods are presented for automatically constructing %low-dimensional coarse spaces of low dimension for domain decomposition algorithms. These constructions use equivalence classes of nodes on the interface between the subdomains into which the domain of a given elliptic problem has been subdivided, e.g., by a mesh partitioner such as METIS; these equivalence classes already play a central role in the design, analysis, and programming of many domain decomposition algorithms. The coarse space elements are well defined even for irregular subdomains, are continuous, and well suited for use in two-level or multi-level preconditioners such as overlapping Schwarz algorithms. An analysis for scalar elliptic and linear elasticity problems ms reveals that significant reductions in the coarse space dimension can be achieved while not sacrificing the favorable condition number estimates for larger coarse spaces previously developed. These estimates depend primarily on the Lipschitz parameters of the subdomains. Numerical examples for problems in three dimensions are presented to illustrate the methods and to confirm the analysis. In some of the experiments, the coefficients have large discontinuities across the interface between the subdomains, and in some, the subdomains are generated by mesh partitioners.
Title: Random Growth Models
Candidate: Florescu, Laura
Advisor(s): Spencer, Joel
Abstract:
This work explores variations of randomness in networks, and more specifically, how drastically the dynamics and structure of a network change when a little bit of information is added to "chaos". On one hand, I investigate how much determinism in diffusions de-randomizes the process, and on the other hand, I look at how superposing "planted" information on a random network changes its structure in such a way that the "planted" structure can be recovered.
The first part of the dissertation is concerned with rotor-router walks, a deterministic counterpart to random walk, which is the mathematical model of a path consisting of a succession of random steps. I study and show results on the volume (``the range") of the territory explored by the random rotor-router model, confirming an old prediction of physicists.
The second major part in the dissertation consists of two constrained diffusion problems. The questions in this model are to understand the long-term behavior of the models, as well as how the boundary of the processes evolves in time.
The third part is detecting communities in, or more generally, clustering networks. This is a fundamental problem in mathematics, machine learning, biology and economics, both for its theoretical foundations as well as for its practical implications. This problem can be viewed as "planting" some structure in a random network; for example, in cryptography, a code can be viewed as hiding some integers in a random sequence. For such a model with two communities, I show both information theoretic thresholds when it is impossible to recover the communities based on the density of the edges "planted" between the communities, as well as thresholds for when it is computationally possible to recover the communities.
Title: Decision Procedures for Finite Sets with Cardinality, and Local Theories Extensions
Candidate: Bansal, Kshitij
Advisor(s): Barrett, Clark; Wies, Thomas
Abstract:
Many tasks in design, verification, and testing of hardware and computer systems can be reduced to checking satisfiability of logical formulas. Certain fragments of first-order logic that model the semantics of prevalent data types, and hardware and software constructs, such as integers, bit-vectors, and arrays are thus of most interest. The appeal of satisfiability modulo theories (SMT) solvers is that they implement decision procedures for efficiently reasoning about formulas in these fragments. Thus, they can often be used off-the-shelf as automated back-end solvers in verification tools. In this thesis, we expand the scope of SMT solvers by developing decision procedures for new theories of interest in reasoning about hardware and software.
First, we consider the theory of finite sets with cardinality. Sets are a common high-level data structure used in programming; thus, such a theory is useful for modeling program constructs directly. More importantly, sets are a basic construct of mathematics and thus natural to use when mathematically defining the properties of a computer system. We extend a calculus for finite sets to reason about cardinality constraints. The reasoning for cardinality involves tracking how different sets overlap. For an efficient procedure in an SMT solver, we'd like to avoid considering Venn regions directly, which has been the approach in earlier work. We develop a novel technique wherein potentially overlapping regions are considered incrementally. We use a graph to track the interaction of the different regions. Additionally, our technique leverages the procedure for reasoning about the other set operations (besides cardinality) in a modular fashion.
Second, a limitation frequently encountered is that verification problems are often not fully expressible in the theories supported natively by the solvers. Many solvers allow the specification of application-specific theories as quantified axioms, but their handling is incomplete outside of narrow special cases. We show how SMT solvers can be used to obtain complete decision procedures for local theory extensions, an important class of theories that are decidable using finite instantiation of axioms. We present an algorithm that uses E-matching to generate instances incrementally during the search, significantly reducing the number of generated instances compared to eager instantiation strategies.
Title: Detecting Missing and Spurious Edges in Large, Dense Networks Using Parallel Computing
Author(s): Coolidge, Sam; Simon, Dan; Shasha, Dennis
Abstract:
Certain pairs of drugs can cause death from their interaction. Knowledge of such interactions is held in drug interaction networks. The problem is that such networks may miss interactions that should be present and may include interactions that should be absent. Clearly, such information is valuable. Drug interaction networks are not unique in this regard. The same holds for protein-protein interaction networks, ecological networks, and many others. Improving the quality of such networks often requires a ground truth analysis (e.g. more experiments) but Roger Guimerá, Marta Sales-Prado, and their colleagues have shown in several papers that a structural analysis of networks can lead to predictions of missing and spurious edges that can improve those networks. Our contribution in this paper and the accompanying software is to create a program implementing their algorithmic ideas that is parallelizable and easy to modify for researchers who wish to try out new ideas. Our software can be found at https://github.com/samcoolidge/network.
Title: Analyzing Source Code Across Static Conditionals
Candidate: Gazzillo, Paul
Advisor(s): Wies, Thomas
Abstract:
We need better tools for C, such as source browsers, bug finders, and automated refactorings. The problem is that large C systems such as Linux are software product lines, containing thousands of configuration variables controlling every aspect of the software from architecture features to file systems and drivers. The challenge of such configurability is how do software tools accurately analyze all configurations of the source without the exponential explosion of trying them all separately. To this end, we focus on two key subproblems, parsing and the build system. The contributions of this thesis are the following: (1) a configuration-preserving preprocessor and parser called SuperC that preserves configurations in its output syntax tree; (2) a configuration-preserving Makefile evaluator called Kmax that collections Linux's compilation units and their configurations; and (3) a framework for configuration-aware analyses of source code using these tools.
C tools need to process two languages: C itself and the preprocessor. The latter improves expressivity through file includes, macros, and static conditionals. But it operates only on tokens, making it hard to even parse both languages. SuperC is a complete, performant solution to parsing all of C. First, a configuration-preserving preprocessor resolves includes and macros yet leaves static conditionals intact, thus preserving a program's variability. To ensure completeness, we analyze all interactions between preprocessor features and identify techniques for correctly handling them. Second, a configuration-preserving parser generates a well-formed AST with static choice nodes for conditionals. It forks new subparsers when encountering static conditionals and merges them again after the conditionals. To ensure performance, we present a simple algorithm for table-driven Fork-Merge LR parsing and four novel optimizations. We demonstrate SuperC's effectiveness on the x86 Linux kernel.
Large-scale C codebases like Linux are software product families, with complex build systems that tailor the software with myriad features. Such variability management is a challenge for tools, because they need awareness of variability to process all software product lines within the family. With over 14,000 features, processing all of Linux's product lines is infeasible by brute force, and current solutions employ incomplete heuristics. But having the complete set of compilation units with precise variability information is key to static tools such a bug-finders, which could miss critical bugs, and refactoring tools, since behavior-preservation requires a complete view of the software project. Kmax is a new tool for the Linux build system that extracts all compilation units with precise variability information. It processes build system files with a variability-aware \texttt{make} evaluator that stores variables in a conditional symbol table and hoists conditionals around complete statements, while tracking variability information as presence conditions. Kmax is evaluated empirically for correctness and completeness on the Linux kernel. Kmax is compared to previous work for correctness and running time, demonstrating that a complete solution's added complexity incurs only minor latency compared to the incomplete heuristic solutions.
SuperC's configuration-preserving parsing of compilation units and Kmax's project-wide capabilities are in a unique position to process source code across all configurations. Bug-finding is one area where such capability is useful. Bugs may appear in untested combinations of configurations and testing each configuration one-at-a-time is infeasible. For example, one compilation units that defines a global function called by other compilation units may not be linked into the final program due to configuration variable selection. Such a bug can be found with Kmax and SuperC's cross-configuration capability. Cilantro is a framework for creating variability-aware bug-checkers. Kmax is used to determine the complete set of compilation units and the combinations of features that activate them, while SuperC's parsing framework is extended with semantic actions in order implement the checkers. A checker for linker errors across all compilation units in the Linux kernel demonstrates each part of the Cilantro framework and is evaluated on the Linux source code.
Title: Semi-Supervised Learning for Electronic Phenotyping in Support of Precision Medicine
Candidate: Halpern, Yonatan
Advisor(s): Sontag, David
Abstract:
Medical informatics plays an important role in precision medicine, delivering the right information to the right person, at the right time. With the introduction and widespread adoption of electronic medical records, in the United States and world-wide, there is now a tremendous amount of health data available for analysis. Electronic record phenotyping refers to the task of determining, from an electronic medical record entry, a concise descriptor of the patient, comprising of their medical history, current problems, presentation, etc. In inferring such a phenotype descriptor from the record, a computer, in a sense, "understands", the relevant parts of the record. These phenotypes can then be used in downstream applications such as cohort selection for retrospective studies, real-time clinical decision support, contextual displays, intelligent search, and precise alerting mechanisms.
To handle the incomplete data present in medical records, we use the formal framework of probabilistic graphical models with latent or unobserved variables. The first part of the thesis presents two different structural conditions under which learning with latent variables is computationally tractable. The first is the "anchored" condition, where every latent variable has at least one child that is not shared by any other parent. The second is the "singly-coupled" condition, where every latent variable is connected to at least three children that satisfy conditional independence (possibly after a transformation of the data). Variables that satisfy these conditions can be specified by an expert without requiring that the entire structure or its parameters be specified, allowing for effective use of human expertise and making room for statistical learning to do some of the heavy lifting in model learning. For both the anchored and singly-coupled conditions, practical algorithms are presented.
The second part of the thesis describes real-life applications using the anchored condition for electronic phenotyping. A human-in-the-loop learning system and a functioning emergency informatics system for real-time extraction of important clinical variables are described and evaluated.
The algorithms and discussion presented here were developed for the purpose of improving healthcare, but are much more widely applicable, dealing with the very basic questions of identifiability and learning models with latent variables - a problem that lies at the very heart of the natural and social sciences.
Title: Finding Prospects for Shopping Centers: a machine learning approach
Author(s): Kogan, Jonathan; Jain, Rishabh; Jean, Joe; Lowrance, Roy; Shasha, Dennis
Abstract:
We have developed an algorithm that predicts which store types are the best prospects to fill vacancies in shopping centers given the combinations of stores already there. The model is able to make predictions with accuracies up to 81.62% for the first prediction, 90.05% for the first two predictions, 93.34% for the first three predictions, 95.52% for the first four predictions, and 96.48% for the first five predictions. The p-values with respect to a naïve strategy of choosing the store types that are simply most frequent are all below 0.0001%. This paper explains how the system was built and some user tests, not all of which were positive. The system can be found at http://linserv2.cims.nyu.edu:54321. The code for the project can be found at https://github.com/jgk99/Store-Prospector.
Title: Improving Knowledge Base Population with Information Extraction
Candidate: Li, Xiang
Advisor(s): Grishman, Ralph
Abstract:
Knowledge Bases (KBs) are data resources that encode world knowledge in machine-readable formats. Knowledge Base Population (KBP) aims at understanding this knowledge and extending KBs with more semantic information, which is a fundamental problem in Artificial Intelligence. It can benefit a wide range of tasks, such as semantic search and question answering. Information Extraction (IE), the task of discovering important types of facts (entities, relations and events) in unstructured text, is necessary and crucial for successfully populating knowledge bases. This dissertation focuses on four essential aspects of knowledge base population by leveraging IE techniques: extracting facts from unstructured data, validating the extracted information, accelerating and enhancing systems with less annotation effort, and utilizing knowledge bases to improve real-world applications.
First, we investigate the Slot Filling task, which is a key component for knowledge base population. Slot filling aims to collect information from a large collection of news, web, or other sources of documents to determine a set of predefined attributes ("slots") for given person and organization entities. We introduce a statistical language understanding approach to automatically construct personal (user-centric) knowledge bases from conversational dialogs.
Second, we consider how to probabilistically estimate the correctness of the extracted slot values. Despite the significant progress of KBP research and systems in recent years, slot filling approaches are still far from completely reliable. Using the NIST KBP Slot Filling task as a case study, we propose a confidence estimation model based on the Maximum Entropy framework, and demonstrate the effectiveness of this model in both precision and the capability to improve the slot filling aggregation through a weighted voting strategy.
Third, we study rich annotation guided learning to fill the gap between an expert annotator and a feature engineer. We develop an algorithm to enrich features with the guidance of all levels of rich annotations from human annotators. We also evaluate the comparative efficacy, generality and scalability of this framework by conducting case studies on three distinct applications in various domains, including facilitating KBP slot filling systems. Empirical studies demonstrate that with little additional annotation time, we can significantly improve the performance for all tasks.
Finally, we explore utilizing knowledge bases in a real-world application - personalized content recommendation. Traditional systems infer user interests from surface-level features derived from online activity logs and user demographic profiles, rather than deeply understanding the context semantics. We conduct a systematic study to show the effectiveness of incorporating deep semantic knowledge encoded in the entities on modeling user interests, by utilizing the abundance of entity information from knowledge bases.
Title: Improving SAT Solvers by Exploiting Empirical Characteristics of CDCL
Candidate: Oh, Chanseok
Advisor(s): Wies, Thomas
Abstract:
The Boolean Satisfiability Problem (SAT) is a canonical decision problem originally shown to be NP-complete in Cook’s seminal work on the theory of computational complexity. The SAT problem is one of several computational tasks identified by researchers as core problems in computer science. The existence of an efficient decision procedure for SAT would imply P = NP. However, numerous algorithms and techniques for solving the SAT problem have been proposed in various forms in practical settings. Highly efficient solvers are now actively being used, either directly or as a core engine of a larger system, to solve real-world problems that arise from many application domains. These state-of-the-art solvers use the Davis-Putnam-Logemann-Loveland (DPLL) algorithm extended with ConflictDriven Clause Learning (CDCL). Due to the practical importance of SAT, building a fast SAT solver can have a huge impact on current and prospective applications. The ultimate contribution of this thesis is improving the state of the art of CDCL by understanding and exploiting the empirical characteristics of how CDCL works on real-world problems. The first part of the thesis shows empirically that most of the unsatisfiable real-world problems solvable by CDCL have a refutation proof with near-constant width for the great portion of the proof. Based on this observation, the thesis provides an unconventional perspective that CDCL solvers can solve real-world problems very efficiently and often more efficiently just by maintaining a small set of certain classes of learned clauses. The next part of the thesis focuses on understanding the inherently different natures of satisfiable and unsatisfiable problems and their implications on the empirical workings of CDCL. We examine the varying degree of roles and effects of crucial elements of CDCL based on the satisfiability status of a problem. Ultimately, we propose effective techniques to exploit the new insights about the different natures of proving satisfi- ability and unsatisfiability to improve the state of the art of CDCL. In the last part of the thesis, we present a reference solver that incorporates all the techniques described in the thesis. The design of the presented solver emphasizes minimality in implementation while guaranteeing state-of-the-art performance. Several versions of the reference solver have demonstrated top-notch performance, earning several medals in the annual SAT competitive events. The minimal spirit of the reference solver shows that a simple CDCL framework alone can still be made competitive with state-of-the-art solvers that implement sophisticated techniques outside the CDCL framework.
Title: Graph-based Approaches to Resolve Entity Ambiguity
Candidate: Pershina, Maria
Advisor(s): Grishman, Ralph
Abstract:
Information Extraction is the task of automatically extracting structured information from unstructured or semi-structured machine-readable documents. One of the challenges of Information Extraction is to resolve ambiguity between entities either in a knowledge base or in text documents. There are many variations of this problem and it is known under different names, such as coreference resolution, entity disambiguation, entity linking, entity matching, etc. For example, the task of coreference resolution decides whether two expressions refer to the same entity; entity disambiguation determines how to map an entity mention to an appropriate entity in a knowledge base (KB); the main focus of entity linking is to infer that two entity mentions in a document(s) refer to the same real world entity even if they do not appear in a KB; entity matching (also record deduplication, entity resolution, reference reconciliation) is to merge records from databases if they refer to the same object.
Resolving ambiguity and finding proper matches between entities is an important step for many downstream applications, such as data integration, question answering, relation extraction, etc. The Internet has enabled the creation of a growing number of large-scale knowledge bases in a variety of domains, posing a scalability challenge for Information Extraction systems. Tools for automatically aligning these knowledge bases would make it possible to unify many sources of structured knowledge and to answer complex queries. However the efficient alignment of large-scale knowledge bases still poses a considerable challenge.
Various aspects and different settings to resolve ambiguity between entities are studied in this dissertation. A new scalable domain-independent graph-based approach utilizing Personalized Page Rank is developed for entity matching across large-scale knowledge bases and evaluated on datasets of 110 million and 203 million entities. A new model for entity disambiguation between a document and a knowledge base utilizing a document graph and effectively filtering out noise is proposed. A new technique based on a paraphrase detection model is proposed to recognize name variations for an entity in a document. A new approach integrating a graph-based entity disambiguation model and this technique is presented for an entity linking task and is evaluated on a dataset for Â the Text Analysis Conference Entity Discovery and Linking 2014 task.
Title: On the Solution of Elliptic Partial Differential Equations on Regions with Corners II: Detailed Analysis
Author(s): Serkh, Kirill
Abstract:
In this report we investigate the solution of boundary value problems on polygonal domains for elliptic partial differential equations.
Title: Alphacodes: Usable, Secure Transactions with Untrusted Providers using Human Computable Puzzles
Author(s): Sharma, Ashlesh; Chandrasekaran, Varun; Amjad, Fareeha; Shasha, Dennis; Subramanian, Lakshminarayanan
Abstract:
Many banking and commerce payment systems, especially in developing regions, continue to require users to share private or sensitive information in clear-text with untrusted providers exposing them to different forms of man-in-the-middle attacks. In this paper, we introduce Alphacodes, a new paradigm that provides a usable security solution that enables users to perform secure transactions with untrusted parties using the notion of visual puzzles. Alphacodes are designed as verification codes for short message transactions and provide easy authentication of critical portions of a transaction. We describe how Alphacodes can be applied in different use cases and also show two simple applications that we have built using the Alphacodes framework. We show security vulnerabilities in existing systems and show how our protocol overcomes them. We also demonstrate the ease of use of Alphacodes with minimal training using two simple mechanical turk studies. Using another simple real world user study involving 10 users who speak Kannada (local Indian language), we show that the Alphacodes concept can be easily extended to other languages beyond English.
Title: Partition Memory Models for Program Analysis
Candidate: Wang, Wei
Advisor(s): Barrett, Clark; Wies, Thomas
Abstract:
Scalability is a key challenge in static program analyses based on solvers for Satisfiability Modulo Theories (SMT). For imperative languages like C, the approach taken for modeling memory can play a significant role in scalability. The main theme of this thesis is using partitioned memory models to divide up memory based on the alias information derived from a points-to analysis.
First, a general analysis framework based on memory partitioning is presented. It incorporates a points-to analysis as a preprocessing step to determine a conservative approximation of which areas of memory may alias or overlap and splits the memory into distinct arrays for each of these areas.
Then we propose a new cell-based field-sensitive points-to analysis, which is an extension of Steensgaard’s unification-based algorithms. A cell is a unit of access with scalar or record type. Arrays and dynamically memory allocations are viewed as a collection of cells. We show how our points-to analysis yields more precise alias information for programs with complex heap data structures.
Our work is implemented in Cascade, a static analysis framework for C programs. It replaces the former flat memory model that models the memory as a single array of bytes. We show that the partitioned memory models achieve better scalability within Cascade, and the cell-based memory model, in particular, improves the performance significantly, making Cascade a state-of-the-art C analyzer.
Title: Scaling Multicore Databases via Constrained Parallel Execution
Author(s): Wang, Zhaoguo; Mu, Shuai; Cui, Yang; Yi, Han; Chen, Haibo; Li, Jinyang
Abstract:
Multicore in-memory databases often rely on traditional concurrency control schemes such as two-phase-locking (2PL) or optimistic concurrency control (OCC). Unfortunately, when the workload exhibits a non-trivial amount of contention, both 2PL and OCC sacrifice much parallel execution opportunity. In this paper, we describe a new concurrency control scheme, interleaving constrained concurrency control (IC3), which provides serializability while allowing for parallel execution of certain conflicting transactions. IC3 combines the static analysis of the transaction workload with runtime techniques that track and enforce dependencies among concurrent transactions. The use of static analysis simplifies IC3’s runtime design, allowing it to scale to many cores. Evaluations on a 64-core machine using the TPC-C benchmark show that IC3 outperforms traditional concurrency control schemes under contention. It achieves the throughput of 434K transactions/sec on the TPC-C benchmark configured with only one warehouse. It also scales better than several recent concurrent control schemes that also target contended workloads.
Title: A New Strongly Polynomial Algorithm for Computing Fisher Market Equilibria with Spending Constraint Utilities
Candidate: Wang, Zi
Advisor(s): Cole, Richard
Abstract:
This thesis develops and analyzes an algorithm to compute equilibrium prices for a Fisher market in which the buyer utilities are given by spending constraint functions, utility functions originally defined by Devanur and Vazirani.
Vazirani gave a weakly polynomial time algorithm to compute the equilibrium prices. More recently Vegh gave a strongly polynomial algorithm. Here we provide another strongly polynomial algorithm, which arguably is conceptually simpler, although the running time is not always better.
Title: Learning Algorithms from Data
Candidate: Zaremba, Wojciech
Advisor(s): Fergus, Rob; LeCun, Yann
Abstract:
Statistical machine learning is concerned with learning models that describe observations. We train our models from data on tasks like machine translation or object recognition because we cannot explicitly write down programs to solve such problems. A statistical model is only useful when it generalizes to unseen data. Solomonoff has proved that one should choose the model that agrees with the observed data, while preferring the model that can be compressed the most, because such a choice guarantees the best possible generalization. The size of the best possible compression of the model is called the Kolmogorov complexity of the model. We define an algorithm as a function with small Kolmogorov complexity.
This Ph.D. thesis outlines the problem of learning algorithms from data and shows several partial solutions to it. Our data model is mainly neural networks as they have proven to be successful in various domains like object recognition, language modeling, speech recognition and others. First, we examine empirical trainability limits for classical neural networks. Then, we extend them by providing interfaces, which provide a way to read memory, access the input, and postpone predictions. The model learns how to use them with reinforcement learning techniques like Reinforce and Q-learning. Next, we examine whether contemporary algorithms such as convolution layer can be automatically rediscovered. We show that it is possible indeed to learn convolution as a special case in a broader range of models. Finally, we investigate whether it is directly possible to enumerate short programs and find a solution to a given problem. This follows the original line of thought behind the Solomonoff induction. Our approach is to learn a prior over programs such that we can explore them efficiently.
Title: Distributed Stochastic Optimization for Deep Learning
Candidate: Zhang, Sixin
Advisor(s): LeCun, Yann
Abstract:
We study the problem of how to distribute the training of large-scale deep learning models in the parallel computing environment. We propose a new distributed stochastic optimization method called Elastic Averaging SGD (EASGD). We analyze the convergence rate of the EASGD method in the synchronous scenario and compare its stability condition with the existing ADMM method in the round-robin scheme. An asynchronous and momentum variant of the EASGD method is applied to train deep convolutional neural networks for image classification on the CIFAR and ImageNet datasets. Our approach accelerates the training and furthermore achieves better test accuracy. It also requires a much smaller amount of communication than other common baseline approaches such as the DOWNPOUR method.
We then investigate the limit in speedup of the initial and the asymptotic phase of the mini-batch SGD, the momentum SGD, and the EASGD methods. We find that the spread of the input data distribution has a big impact on their initial convergence rate and stability region. We also find a surprising connection between the momentum SGD and the EASGD method with a negative moving average rate. A non-convex case is also studied to understand when EASGD can get trapped by a saddle point.
Finally, we scale up the EASGD method by using a tree structured network topology. We show empirically its advantage and challenge. We also establish a connection between the EASGD and the DOWNPOUR method with the classical Jacobi and the Gauss-Seidel method, thus unifying a class of distributed stochastic optimization methods.
Title: Pushing the Limits of Additive Fabrication Technologies
Candidate: Zhou, Qingnan (James)
Advisor(s): Zorin, Denis
Abstract:
A rough symmetry can be observed in the stock price of 3D Systems (NYSE:DDD), the leading and largest 3D printer manufacturer, from its IPO on June 3, 2011 to the beginning of 2016. The price sky rocketed nearly 600% from 2011 to the end of 2013, and took a free fall back to its original value by 2016. Coincidentally, it is also the period during which I got my hands dirty and investigated some of the toughest challenges as well as exciting new possibilities associated with different types of 3D printing technologies. In this thesis, I documented my attempts from 3 different angles to push the limits of 3D printing: printability, microstructure design and robust geometry processing with mesh arrangements.
Printability check has long been the bottleneck that prevents 3D printing from scaling up. Oftentimes, designers of 3D models lack the expertise or tools to ensure 3D printability. 3D printing service providers typically rely human inspections to filter out unprintable designs. This process is manual and error-prone. As designs become ever more complex, manual printability check becomes increasingly difficult. To tackle this problem, my colleagues and I proposed an algorithm to automatically determine structurally weak regions and the worst-case usage scenario to break a given model. We validate the algorithm by physically break a number of real 3D printed designs.
A key distinctive feature of 3D printing technologies is that the cost and time of fabrication is uncorrelated with geometric complexity. This opens up many exciting new possibilities. In particular, by pushing geometric complexity to the extreme, 3D printing has the potential of fabricating soft, deformable shapes with microscopic structures using a single raw material. In our recent SIGGRAPH publication, my colleagues and I have not only demonstrated fabricating microscopic frame structures is possible but also proposed an entire pipeline for designing spatially varying microstructures to satisfy target material properties or deformation goals.
With the boost of 3D printing technologies, 3D models have become more abundant and easily accessible than ever before. These models are sometimes known as "wild" models because they differ significantly in complexity and quality from traditional models in graphics researches. This poses a serious challenge in robustly analyzing 3D designs. Many state-of-the-art geometry processing algorithms/libraries are ill-prepared for dealing with "wild" models that are non-manifold, self-intersecting, locally degenerate and/or containing multiple and possibly nested components. In our most recent SIGGRAPH submission, we proposed a systematic recipe based on mesh arrangements for conducting a family of exact constructive solid geometry operations. We exhaustively tested our algorithm on 10,000 "wild" models crawled from Thingiverse, a popular online shape repository. Both the code and the dataset are freely available to the public.
Title: Adaptive Selection of Primal Constraints for Isogeometric BDDC Deluxe Preconditioners
Author(s): Beirão da Veiga, L.; Pavarino, L. F.; Scacchi, S.; Widlund, O. B.; Zampini, S.
Abstract:
Isogeometric analysis has been introduced as an alternative to finite element methods in order to simplify the integration of CAD software and the discretization of variational problems of continuum mechanics. In contrast with the finite element case, the basis functions of isogeometric analysis are often not nodal. As a consequence, there are fat interfaces which can easily lead to an increase in the number of interface variables after a decomposition of the parameter space into subdomains. Building on earlier work on the deluxe version of the BDDC family of domain decomposition algorithms, several adaptive algorithms are here developed for scalar elliptic problems in an effort to decrease the dimension of the global, coarse component of these preconditioners. Numerical experiments provide evidence that this work can be successful, yielding scalable and quasi-optimal adaptive BDDC algorithms for isogeometric discretizations.
Title: An Adaptive Choice of Primal Constraints for BDDC Domain Decomposition Algorithms
Author(s): Calvo, Juan G.; Widlund, Olof B.
Abstract:
An adaptive choice for primal spaces, based on parallel sums, is developed for BDDC deluxe methods and elliptic problems in three dimensions. The primal space, which form the global, coarse part of the domain decomposition algorithm, and which is always required for any competitive algorithm, is defined in terms of generalized eigenvalue problems related to subdomain edges and faces; selected eigenvectors associated to the smallest eigenvalues are used to enhance the primal spaces. This selection can be made automatic by using tolerance parameters specified for the subdomain faces and edges. Numerical results verify the results and provide a comparison with primal spaces commonly used. They include results for cubic subdomains as well as subdomains obtained by a mesh partitioner. Different distributions for the coefficients are also considered, with constant coefficients, highly random values, and channel distributions.
Title: Domain Decomposition Methods for Problems in H(curl)
Author(s): Calvo, Juan Gabriel
Abstract:
Two domain decomposition methods for solving vector field problems posed in H(curl) and discretized with Nédélec finite elements are considered. These finite elements are conforming in H(curl).
A two-level overlapping Schwarz algorithm in two dimensions is analyzed, where the subdomains are only assumed to be uniform in the sense of Peter Jones. The coarse space is based on energy minimization and its dimension equals the number of interior subdomain edges. Local direct solvers are based on the overlapping subdomains. The bound for the condition number depends only on a few geometric parameters of the decomposition. This bound is independent of jumps in the coefficients across the interface between the subdomains for most of the different cases considered.
A bound is also obtained for the condition number of a balancing domain decomposition by constraints (BDDC) algorithm in two dimensions, with Jones subdomains. For the primal variable space, a continuity constraint for the tangential average over each interior subdomain edge is imposed. For the averaging operator, a new technique named deluxe scaling is used. The optimal bound is independent of jumps in the coefficients across the interface between the subdomains.
Furthermore, a new coarse function for problems in three dimensions is introduced, with only one degree of freedom per subdomain edge. In all the cases, it is established that the algorithms are scalable. Numerical results that verify the results are provided, including some with subdomains with fractal edges and others obtained by a mesh partitioner.
Title: Big Data Analytics for Development: Events, Knowledge Graphs and Predictive Models
Candidate: Chakraborty, Sunandan
Advisor(s): Subramanian, Lakshminarayanan; Nyarko, Yaw
Abstract:
Volatility in critical socio-economic indices can have a significant negative impact on global development. This thesis presents a suite of novel big data analytics algorithms that operate on unstructured Web data streams to automatically infer events, knowledge graphs and predictive models to understand, characterize and predict the volatility of socioeconomic indices.
This thesis makes four important research contributions. First, given a large volume of diverse unstructured news streams, we present new models for capturing events and learning spatio-temporal characteristics of events from news streams. We specifically explore two types of event models in this thesis: one centered around the concept of event triggers and a probabilistic meta-event model that explicitly delineates named entities from text streams to learn a generic class of meta-events. The second contribution focuses on learning several different types of knowledge graphs from news streams and events: a) Spatio-temporal article graphs capture intrinsic relationships between different news articles; b) Event graphs characterize relationships between events and given a news query, provide a succinct summary of a timeline of events relating to a query; c) Event-phenomenon graphs that provide a condensed representation of classes of events that relate to a given phenomena at a given location and time; d) Causality testing on word-word graphs which can capture strong spatio-temporal relationships between word occurrences in news streams; e) Concept graphs that capture relationships between different word concepts that occur in a given text stream.
The third contribution focuses on connecting the different knowledge graph representations and structured time series data corresponding to a socio-economic index to automatically learn event-driven predictive models for the given socio-economic index to predict future volatility. We propose several types of predictive models centered around our two event models: event triggers and probabilistic meta-events. The final contribution focuses on a broad spectrum of inference case studies for different types of socio-economic indices including food prices, stock prices, disease outbreaks and interest rates. Across all these indices, we show that event-driven predictive models provide significant improvements in prediction accuracy over state-of-the-art techniques.
Title: SMT-Based and Disjunctive Relational Abstract Domains for StaticAnalysis
Candidate: Chen, Junjie
Advisor(s): Patrick Cousot
Abstract:
Abstract Interpretation is a theory of sound approximation of program semantics. In recent decades, it has been widely and successfully applied to the static analysis of computer programs. In this thesis, we will work on abstract domains, one of the key concepts in abstract interpretation, which aim at automatically collecting information about the set of all possible values of the program variables. We will focus, in particularly, on two aspects: the combination with theorem provers and the refinement of existing abstract domains.
Satisfiability modulo theories (SMT) solvers are popular theorem provers, which proved to be very powerful tools for checking the satisfiability of first-order logical formulas with respect to some background theories. In the first part of this thesis, we introduce two abstract domains whose elements are logical formulas involving finite conjunctions of affine equalities and finite conjunctions of linear inequalities. These two abstract domains rely on SMT solvers for the computation of transformations and other logical operations.
In the second part of this thesis, we present an abstract domain functor whose elements are binary decision trees. It is parameterized by decision nodes which are a set of boolean tests appearing in the programs and by a numerical or symbolic abstract domain whose elements are the leaves. This new binary decision tree abstract domain functor provides a flexible way of adjusting the cost/precision ratio in path-dependent static analysis.
Title: Iris: Mitigating Phase Noise in Millimeter Wave OFDM Systems
Candidate: Dhananjay, Aditya
Advisor(s): Li, Jinyang
Abstract:
Next-generation wireless networks are widely expected to operate over millimeter-wave (mmW) frequencies of over 28GHz. These bands mitigate the acute spectrum shortage in the conventional microwave bands of less than 6GHz. The shorter wavelengths in these bands also allow for building dense antenna arrays on a single chip, thereby enabling various MIMO configurations and highly directional links that can increase the spatial reuse of spectrum.
While attempting to build a practical over-the-air (OTA) link over mmW, we realized that the traditional baseband processing techniques used in the microwave bands simply could not cope with the exacerbated frequency offsets (or phase noise) observed in the RF oscillators at these bands. While the frequency offsets are large, the real difficulty arose from the fact that they varied significantly over very short time-scales.Traditional feedback loop techniques still left significant residual offsets, which in turn led to inter-carrier-interference (ICI). The result was very high symbol error rates (SER).
This thesis presents Iris, a baseband processing block that enables clean mmW links, even in the presence of previously fatal amounts of phase noise. Over real mmW hardware, Iris reduces the SER by one to two orders of magnitude, as compared to competing techniques.
Title: Predicting Images using Convolutional Networks: Visual Scene Understanding with Pixel Maps
Candidate: Eigen, David
Advisor(s): Fergus, Rob
Abstract:
In the greater part of this thesis, we develop a set of convolutional networks that infer predictions at each pixel of an input image. This is a common problem that arises in many computer vision applications: For example, predicting a semantic label at each pixel describes not only the image content, but also fine-grained locations and segmenta- tions; at the same time, finding depth or surface normals provide 3D geometric relations between points. The second part of this thesis investigates convolutional models also in the contexts of classification and unsupervised learning.
To address our main objective, we develop a versatile Multi-Scale Convolutional Network that can be applied to diverse vision problems using simple adaptations, and apply it to predict depth at each pixel, surface normals and semantic labels. Our model uses a series of convolutional network stacks applied at progressively finer scales. The first uses the entire image field of view to predict a spatially coarse set of feature maps based on global relations; subsequent scales correct and refine the output, yielding a high resolution prediction. We look exclusively at depth prediction first, then generalize our method to multiple tasks. Our system achieves state-of-the-art results on all tasks we investigate, and can match many image details without the need for superpixelation.
Leading to our multi-scale network, we also design a purely local convolutional network to remove dirt and raindrops present on a window surface, which learns to identify and inpaint compact corruptions. We also we investigate a weighted nearest-neighbors labeling system applied to superpixels, in which we learn weights for each example, and use local context to find rare class instances.
In addition, we investigate the relative importance of sizing parameters using a recursive convolutional network, finding that network depth is most critical. We also develop a Convolutional LISTA Autoencoder, which learns features similar to stacked sparse coding at a fraction of the cost, combine it with a local entropy objective, and describe a convolutional adaptation of ZCA whitening.
Title: Kmax: Analyzing the Linux Build System
Author(s): Gazzillo, Paul
Abstract:
Large-scale C software like Linux needs software engineering tools. But such codebases are software product families, with complex build systems that tailor the software with myriad features. This variability management is a challenge for tools, because they need awareness of variability to process all software product lines within the family. With over 14,000 features, processing all of Linux's product lines is infeasible by brute force, and current solutions employ incomplete heuristics. But having the complete set of compilation units with precise variability information is key to static tools such a bug-finders, which could miss critical bugs, and refactoring tools, since behavior-preservation requires a complete view of the software project. Kmax is a new tool for the Linux build system that extracts all compilation units with precise variability information. It processes build system files with a variability-aware make evaluator that stores variables in a conditional symbol table and hoists conditionals around complete statements, while tracking variability information as presence conditions. Kmax is evaluated empirically for correctness and completeness on the Linux kernel. Kmax is compared to previous work for correctness and running time, demonstrating that a complete solution's added complexity incurs only minor latency compared to the incomplete heuristic solutions.
Title: Unsupervised Feature Learning in Computer Vision
Candidate: Goroshin, Ross
Advisor(s): LeCun, Yann
Abstract:
Much of computer vision has been devoted to the question of representation through feature extraction. Ideal features transform raw pixel intensity values to a representation in which common problems such as object identification, tracking, and segmentation are easier to solve. Recently, deep feature hierarchies have proven to be immensely successful at solving many problems in computer vision. In the supervised setting, these hierarchies are trained to solve specific problems by minimizing an objective function of the data and problem specific label information. Recent findings suggest that despite being trained on a specific task, the learned features can be transferred across multiple visual tasks. These findings suggests that there exists a generically useful feature representation for natural visual data.
This work aims to uncover the principles that lead to these generic feature representations in the unsupervised setting, which does not require problem specific label information. We begin by reviewing relevant prior work, particularly the literature on autoencoder networks and energy based learning. We introduce a new regularizer for autoencoders that plays an analogous role to the partition function in probabilistic graphical models. Next we explore the role of specialized encoder architectures for sparse inference. The remainder of the thesis explores visual feature learning from video. We establish a connection between slow-feature learning and metric learning, and experimentally demonstrate that semantically coherent metrics can be learned from natural videos. Finally, we posit that useful features linearize natural image transformations in video. To this end, we introduce a new architecture and loss for training deep feature hierarchies that linearize the transformations observed in unlabeled natural video sequences by learning to predict future frames in the presence of uncertainty.
Title: Efficient and Trustworthy Theory Solver for Bit-vectors in SatisfiabilityModulo Theories
Candidate: Hadarean, Liana
Advisor(s): Barrett, Clark
Abstract:
As software and hardware systems grow in complexity, automated techniques for ensuring their correctness are becoming increasingly important. Many modern formal verification tools rely on back-end satisfiability modulo theories (SMT) solvers to discharge complex verification goals. These goals are usually formalized in one or more fixed first-order logic theories, such as the theory of fixed-width bit-vectors. The theory of bit-vectors offers a natural way of encoding the precise semantics of typical machine operations on binary data. The predominant approach to deciding the bit-vector theory is via eager reduction to propositional logic. While this often works well in practice, it does not scale well as the bit-width and number of operations increase. The first part of this thesis seeks to fill this gap, by exploring efficient techniques of solving bit-vector constraints that leverage the word-level structure. We propose two complementary approaches: an eager approach that takes full advantage of the solving power of off the shelf propositional logic solvers, and a lazy approach that combines on-the-fly algebraic reasoning with efficient propositional logic solvers. In the second part of the thesis, we propose a proof system for encoding automatically checkable refutation proofs in the theory of bit-vectors. These proofs can be automatically generated by the SMT solver, and act as a certificate for the correctness of the result.
Title: A Crop Recommendation Tool for Organic Farmers
Author(s): Hsu, Jasmine; Shasha, Dennis
Abstract:
We describe the data sources and machine learning algorithms that go into the current version of http://www.whatcanifarm.com , a website to help prospective organic farmers determine what to grow given the climate characterized by their zip code.
Title: Predicting the Market Value of Single-Family Residences
Candidate: Lowrance, Roy
Advisor(s): LeCun, Yann; Shasha, Dennis
Abstract:
This work develops the best linear model of residential real estate prices for 2003 through 2009 in Los Angeles County. It differs from other studies comparing models for predicting house prices by covering a larger geographic area than most, more houses than most, a longer time period than most, and the time period both before and after the real estate price boom in the United States.
In addition, it open sources all of the software. We test designs for linear models to determine the best form for the model as well as the training period, features, and regularizer that produce the lowest errors. We compare the best of our linear models to random forests and point to directions for further research.
Title: Building Fast, CPU-Efficient Distributed Systems on Ultra-Low Latency, RDMA-Capable Networks
Candidate: Mitchell, Christopher
Advisor(s): Li, Jinyang
Abstract:
Modern datacenters utilize traditional Ethernet interconnects to connect hundreds or thousands of machines. Although inexpensive and ubiquitous, Ethernet imposes design constraints on datacenter-scale distributed storage systems that use traditional client-server architectures. Recent technological trends indicate that future datacenters will embrace interconnects with ultra-low latency, high bandwidth, and the ability to offload work from servers to clients. Future datacenter-scale distributed storage systems will need to be designed specifically to exploit these features. This thesis explores what these features mean for large-scale in-memory storage systems, and derives two key insights for building RDMA-aware distributed systems.
First, relaxing locality between data and computation is now practical: data can be copied from servers to clients for computation. Second, selectively relaxing data-computation locality makes it possible to optimally balance load between server and client CPUs to maintain low application latency. This thesis presents two in-memory distributed storage systems built around these two insights, Pilaf and Cell, that demonstrate effective use of ultra-low-latency, RDMA-capable interconnects. Through Pilaf and Cell, this thesis demonstrates that by combining RDMA and message passing to selectively relax locality, systems can achieve ultra-low latency and optimal load balancing with modest CPU resources.
Title: BDDC Algorithm with Deluxe Scaling and Adaptive Selection of Primal Constraints for Raviart-Thomas Vector Fields
Author(s): Oh, Duk-Soon; Widlund, Olof B.; Zampini, Stefano; Dohrmann, Clark R.
Abstract:
A BDDC domain decomposition preconditioner is defined by a coarse component, expressed in terms of primal constraints, a weighted average across the interface between the subdomains, and local components given in terms of solvers of local subdomain problems. BDDC methods for vector field problems discretized with Raviart-Thomas finite elements are introduced. The methods are based on a new type of weighted average and an adaptive selection of primal constraints developed to deal with coefficients with high contrast even inside individual subdomains. For problems with very many subdomains, a third level of the preconditioner is introduced.
Under the assumption that the subdomains are all built from elements of a coarse triangulation of the given domain, and that the material parameters are constant in each subdomain, a bound is obtained for the condition number of the preconditioned linear system which is independent of the values and the jumps of the coefficients across the interface and has a polylogarithmic condition number bound in terms of the number of degrees of freedom of the individual subdomains. Numerical experiments, using the PETSc library, and a large parallel computer, for two and three dimensional problems are also presented which support the theory and show the effectiveness of the algorithms even for problems not covered by the theory. Included are also experiments with a variety of finite element approximations.
Title: Practical SMT-Based Type Error Localization
Author(s): Pavlinovic, Zvonimir; Wies, T
Abstract:
Compilers for statically typed functional programming languages are notorious for generating confusing type error messages. When the compiler detects a type error, it typically reports the program location where the type checking failed as the source of the error. Since other error sources are not even considered, the actual root cause is often missed. A more adequate approach is to consider all possible error sources and report the most useful one subject to some usefulness criterion. In our previous work, we showed that this approach can be formulated as an optimization problem related to satisfiability modulo theories (SMT). This formulation cleanly separates the heuristic nature of usefulness criteria from the underlying search problem. Unfortunately, algorithms that search for an optimal error source cannot directly use principal types which are crucial for dealing with the exponential-time complexity of the decision problem of polymorphic type checking. In this paper, we present a new algorithm that efficiently finds an optimal error source in a given ill-typed program. Our algorithm uses an improved SMT encoding to cope with the high complexity of polymorphic typing by iteratively expanding the typing constraints from which principal types are derived. The algorithm preserves the clean separation between the heuristics and the actual search. We have implemented our algorithm for OCaml. In our experimental evaluation, we found that the algorithm reduces the running times for optimal type error localization from minutes to seconds and scales better than previous localization algorithms.
Title: Instance Segmentation of RGBD Scenes
Candidate: Silberman, Nathan
Advisor(s): Fergus, Rob
Abstract:
The vast majority of literature in scene parsing can be described as semantic pixel labeling or semantic segmentation: predicting the semantic class of the object represented by each pixel in the scene. Our familiar perception of the world, however, provides a far richer representation. Firstly, rather than just being able to predict the semantic class of a location in a scene, humans are able to reason about object instances. Discriminating between a region that might represent a single object versus ten objects is a crucial and basic faculty. Secondly, rather than reasoning about objects as merely occupying the space visible from a single vantage point, we are able to quickly and easily reason about an object's true extent in 3D. Thirdly, rather than viewing a scene as a collection of objects independently existing in space, humans exhibit a representation of scenes that is highly grounded through a intuitive model of physics. Such models allow us to reason about how objects relate physically: via physical support relationships.
Instance segmentation is the task of segmenting a scene into regions which correspond to individual object instances. We argue that this task is not only closer to our own perception of the world than semantic segmentation, but also directly allows for subsequent reasoning about a scenes constituent elements. We explore various strategies for instance segmentation in indoor RGBD scenes.
Firstly, we explore tree-based instance segmentation algorithms. The utility of trees for semantic segmentation has been thoroughly demonstrated and we adapt them to instance segmentation and analyze both greedy and global approaches to inference.
Next, we investigate exemplar-based instance segmentation algorithms, in which a set of representative exemplars are chosen from a large pool of regions and pixels are assigned to exemplars. Inference can either be performed in two stages, exemplar selection followed by pixel-to-exemplar assignment, or in a single joint reasoning stage. We consider the advantages and disadvantages of each approach.
We introduce the task of support-relation prediction in which we predict which objects are physically supporting other objects. We propose an algorithm and a new set of features for performing discriminative support prediction, we demonstrate the effectiveness of our method and compare training mechanisms.
Finally, we introduce an algorithm for inferring scene and object extent. We demonstrate how reasoning about 3D extent can be done by extending known 2D methods and highlight the strengths and limitations of this approach.
Title: Localization of Humans in Images Using Convolutional Networks
Candidate: Tompson, Jonathan
Advisor(s): Bregler, Christopher
Abstract:
Tracking of humans in images is a long standing problem in computer vision research for which, despite significant research effort, an adequate solution has not yet emerged. This is largely due to the fact that human body localization is complicated and difficult; potential solutions must find the location of body joints in images with invariance to shape, lighting and texture variation and it must do so in the presence of occlusion and incomplete data. However, despite these significant challenges, this work will present a framework for human body pose localization that not only offers a significant improvement over existing traditional architectures, but has sufficient localization performance and computational efficiency for use in real-world applications.
At it's core, this framework makes use of Convolutional Networks to infer the location of body joints efficiently and accurately. We describe solutions to two applications 1) hand-tracking from a depth image source and 2) human body-tracking from and RGB image source. For both these applications we show that Convolutional Networks are able to significantly out-perform existing state-of-the-art.
We propose a new hybrid architecture that consists of a deep Convolutional Network and a Probabilistic Graphical Model which can exploit structural domain constraints such as geometric relationships between body joint locations to improve tracking performance. We then explore the use of both color and motion features to improve tracking performance. Finally we introduce a novel architecture which includes an efficient ‘position refinement’ model that is trained to estimate the joint offset location within a small region of the image. This refinement model allows our network to improve spatial localization accuracy even with large amounts of spatial pooling.
Title: Acronym Disambiguation
Author(s): Turtel, Benjamin D.; Shasha, Dennis
Abstract:
Acronym disambiguation is the process of determining the correct expansion of an acronym in a given context. We describe a novel approach for expanding acronyms, by identifying acronym / expansion pairs in a large training corpus of text from Wikipedia and using these as a training dataset to expand acronyms based on word frequencies. On instances in which the correct acronym expansion has at least one instance in our training set (therefore making correct expansion possible), and in which the correct expansion is not the only expansion of an acronym seen in our training set (therefore making the expansion decision a non-trivial decision), we achieve an average accuracy of 88.6%. On a second set of experiments using user-submitted documents, we achieve an average accuracy of 81%.
Title: Joint Training of a Neural Network and a Structured Model for Computer Vision
Candidate: Wan, Li
Advisor(s): Fergus, Rob
Abstract:
Identifying objects and telling where they are in real world images is one of the most important problems in Artificial Intelligence. The problem is challenging due to: occluded objects, varying object viewpoints and object deformations. This makes the vision problem extremely difficult and cannot be efficiently solved without learning.
This thesis explores hybrid systems that combine a neural network as a trainable feature extractor and structured models that capture high level information such as object parts. The resulting models combine the strengths of the two approaches: a deep neural network which provides a powerful non-linear feature transformation and a high level structured model which integrates domain-specific knowledge. We develop discriminative training algorithms to jointly optimize these entire models end-to-end.
First, we proposed a unified model which combines a deep neural network with a latent topic model for image classification. The hybrid model is shown to outperform models based solely on neural networks or topic model alone. Next, we investigate techniques for training a neural network system, introducing an effective way of regularizing the network called DropConnect. DropConnect allows us to train large models while avoiding over-fitting. This yields state-of-the-art results on a variety of standard benchmarks for image classification. Third, we worked on object detection for PASCAL challenge. We improved the deformable parts model and proposed a new non-maximal suppression algorithm. This system was the joint winner of the 2011 challenge. Finally, we develop a new hybrid model which integrates a deep network, deformable parts model and non-maximal suppression. Joint training of our hybrid model shows clear advantage over train each component individually, and achieving competitive result on standard benchmarks.
Title: Partition Memory Models in Program Analysis
Candidate: Wang, Wei
Advisor(s): Barrett, Clark
Abstract:
Scalability is a key challenge in static program analyses based on solvers for Satisfiability Modulo Theories (SMT). For imperative languages like C, the approach taken for modeling memory can play a significant role in scalability. The main theme of this thesis is using partitioned memory models to divide up memory based on the alias information derived from a points-to analysis.
First, a general analysis framework based on memory partitioning is presented. It incorporates a points-to analysis as a preprocessing step to determine a conservative approximation of which areas of memory may alias or overlap and splits the memory into distinct arrays for each of these areas.
Then we propose a new cell-based field-sensitive points-to analysis, which is an extension of Steensgaard's unification-based algorithms. A cell is a unit of access with scalar or record type. Arrays and dynamically memory allocations are viewed as a collection of cells. We show how our points-to analysis yields more precise alias information for programs with complex heap data structures.
Our work is implemented in Cascade, a static analysis framework for C programs. It replaces the former at memory model that models the memory as a single array of bytes. We show that the partitioned memory models achieve better scalability within Cascade, and the cell-based memory model, in particular, improves the performance significantly, making Cascade a state-of-the-art C analyzer.
Title: On the Human Form: Efficient acquisition, modeling and manipulation of thehuman body
Candidate: Braga, Otavio
Advisor(s): Geiger, Davi
Abstract:
This thesis concerns the acquisition, modeling and manipulation of the human form.
First, we acquire body models. We introduce an efficient bootstraped algorithm that we employed to register over 2,000 high resolution body scans of male and female adult subjects. Our algorithm outputs not only the traditional vertex correspondences, but also directly produces a high quality model which can be immediately deformed. We then employ the result to fit noisy depth maps coming from now commercially available 3D sensors such as Microsoft's Kinect and PrimeSense's Carmine.
We conclude by describing a new real-time system for image-based body manipulation called BodyJam, that lets you change your outfit with a finger snap. BodyJam is inspired by a technique invented by the surrealists a century ago: "Exquisite corpse", a method by which a collection of images (of body parts) is collectively assembled. BodyJam does it on a video display that mirrors the pose in real-time of a real-person standing in front of the camera/display mirror, and allows the user to change clothes and other appearance attributes. Using Microsoft's Kinect, poses are matched to a video database of different torsos and legs, and "pages" showing different clothes are turned by handwitch focus to the topic of body manipulation. We first revisit the more traditional way of specifying bodies from a set of measurements, such as coming from clothing sizing charts, showing how the statistics of the population learned during the registration can aid us in accurately defining the body shape. We then introduce a new manipulation metaphor, where we navigate through the space of body shapes and poses by directly dragging the body mesh surface.
We conclude by describing a new real-time system for image-based body manipulation called BodyJam, that lets you change your outfit with a finger snap. BodyJam is inspired by a technique invented by the surrealists a century ago: "Exquisite Corpse", a method by which a collection of images (of body parts) is collectively assembled. BodyJam does it on a video display that mirrors the pose in real-time of a real-person standing in front of the camera/display mirror, and allows the user to change clothes and other appearance attributes. Using Microsoft's Kinect, poses are matched to a video database of different torsos and legs, and "pages" showing different clothes are turned by hand gestures.
Title: Overlapping Schwarz Algorithms for Almost Incompressible Linear Elasticity
Author(s): Cai, Mingchao; Pavarino, Luca F.; Widlund, Olof B.
Abstract:
Low order finite element discretizations of the linear elasticity system suffer increasingly from locking effects and ill-conditioning, when the material approaches the incompressible limit, if only the displacement variable are used. Mixed finite elements using both displacement and pressure variables provide a well-known remedy, but they yield larger and indefinite discrete systems for which the design of scalable and efficient iterative solvers is challenging. Two-level overlapping Schwarz preconditioner for the almost incompressible system of linear elasticity, discretized by mixed finite elements with discontinuous pressures, are constructed and analyzed. The preconditioned systems are accelerated either by a GMRES (generalized minimum residual) method applied to the resulting discrete saddle point problem or by a PCG (preconditioned conjugate gradient) method applied to a positive definite, although extremely ill-conditioned, reformulation of the problem obtained by eliminating all pressure variables on the element level. A novel theoretical analysis of the algorithm for the positive definite reformulation is given by extending some earlier results by Dohrmann and Widlund. The main result of the paper is a bound on the condition number of the algorithm which is cubic in the relative overlap and grows logarithmically with the number of elements across individual subdomains but is otherwise independent of the number of subdomains, their diameters and mesh sizes, and the incompressibility of the material and possible discontinuities of the material parameters across the subdomain interfaces. Numerical results in the plane confirm the theory and also indicate that an analogous result should hold for the saddle point formulation, as well as for spectral element discretizations.
Title: A two-level overlapping Schwarz method for H(curl) in two dimensions with irregular subdomains
Author(s): Calvo, Juan G.
Abstract:
A bound is obtained for the condition number of a two-level overlapping Schwarz algorithm for problems posed in H(curl) in two dimensions, where the subdomains are only assumed to be uniform in the sense of Peter Jones. The coarse space is based on energy minimization and its dimension equals the number of interior subdomain edges. Local direct solvers are used on the overlapping subdomains. Our bound depends only on a few geometric parameters of the decomposition. This bound is independent of jumps in the coefficients across the interface between the subdomains for most of the different cases considered. Numerical experiments that verify the result are shown, including some with subdomains with fractal edges and others obtained by a mesh partitioner.
Title: A BDDC algorithm with deluxe scaling for H(curl) in two dimensions with irregular subdomains
Author(s): Calvo, Juan G.
Abstract:
A bound is obtained for the condition number of a BDDC algorithm for problems posed in H(curl) in two dimensions, where the subdomains are only assumed to be uniform in the sense of Peter Jones. For the primal variable space, a continuity constraint for the tangential average over each interior subdomain edge is imposed.
For the averaging operator, a new technique named deluxe scaling is used. Our bound is independent of jumps in the coefficients across the interface between the subdomains and depends only on a few geometric parameters of the decomposition. Numerical results that verify the result are shown, including some with subdomains with fractal edges and others obtained by a mesh partitioner.
Title: Analyzing Tatonnement Dynamics in Economic Markets
Candidate: Cheung, Yun Kuen
Advisor(s): Cole, Richard
Abstract:
The impetus for this dissertation is to explain why well-functioning markets might be able to stay at or near a market equilibrium. We argue that tatonnement, a natural, simple and distributed price update dynamic in economic markets, is a plausible candidate to explain how markets might reach their equilibria.
Tatonnement is broadly defined as follows: if the demand for a good is more than the supply, increase the price of the good, and conversely, decrease the price when the demand is less than the supply. Prior works show that tatonnement converges to market equilibrium in some markets while it fails to converge in other markets. Our goal is to extend the classes of markets in which tatonnement is shown to converge. The prior positive results largely concerned markets with substitute goods. We seek market constraints which enable tatonnement to converge in markets with complementary goods, or with a mixture of substitutes and complementary goods. We also show fast convergence rates for some of these markets.
We introduce an amortized analysis technique to handle asynchronous events - in our case asynchronous price updates. On the other hand, for some markets we show that tatonnement is equivalent to generalized gradient descent (GGD). The amortized analysis and our analysis on GGD may be of independent interests.
Title: A BDDC algorithm with deluxe scaling for three-dimensional H(curl) problems
Author(s): Dohrmann, Clark R.; Widlund, Olof B.
Abstract:
In this paper, we present and analyze a BDDC algorithm for a class of elliptic problems in the three-dimensional H(curl) space. Compared with existing results, our condition number estimate requires fewer assumptions and also involves two fewer powers of log(H/h), making it consistent with optimal estimates for other elliptic problems. Here, H/h is the maximum of H _{ i } /h _{ i } over all subdomains, where H _{ i } and h _{ i } are the diameter and the smallest element diameter for the subdomain Ω _{ i } .
The analysis makes use of two recent developments. The first is a new approach to averaging across the subdomain interfaces, while the second is a new technical tool which allows arguments involving trace classes to be avoided. Numerical examples are presented to confirm the theory and demonstrate the importance of the new averaging approach in certain cases.
Title: Low-latency Image Recognition withGPU-accelerated Convolutional Networksfor Web-based Services
Candidate: Huang, Fu Jie
Advisor(s): LeCun, Yann
Abstract:
In this work, we describe an application of convolutional networks to object classification and detection in images. The task of image based object recognition is surveyed in the first chapter. Its application in internet advertisement is one of the main motivations of this work.
The architecture of the convolutional networks is described in details in the following chapter. Stochastic gradient descent is used to train the networks.
We then describe the data collection and labelling process. The set of training data labelled basically decides what kind of recognizer is being built. Four binary classifers are trained for the object types of sailboat, car, motorbike, and dog.
GPU based massive parallel implementation of the convolutional networks is built. This enables us to run the convolution operations at close to 40 times faster than running on a traditional CPU. Details about how to implement the convolutional operation on NVIDIA GPUs using CUDA is disscused.
In order to apply the object recognizer in a production environment where millions of images are processed daily, we have built a platform with cloud computing. We describe how large scale and low latency image processing can be achieved with such a system.
Title: Effective Algorithms for the Satisfiability of Quantifier-Free Formulas Over Linear Real and Integer Arithmetic
Candidate: King, Tim
Advisor(s): Barrett, Clark
Abstract:
A core technique of modern tools for formally reasoning about computing systems is generating and dispatching queries to automated theorem provers, including Satisfiability Modulo Theories (SMT) provers. SMT provers aim at the tight integration of decision procedures for propositional satisfiability and decision procedures for fixed first-order theories ‒ known as theory solvers. This thesis presents several advancements in the design and implementation of theory solvers for quantifier-free linear real, integer, and mixed integer and real arithmetic. These are implemented within the SMT system CVC4. We begin by formally describing the Satisfiability Modulo Theories problem and the role of theory solvers within CVC4. We discuss known techniques for building solvers for quantifier-free linear real, integer, and mixed integer and real arithmetic around the Simplex for DPLL(T) algorithm. We give several small improvements to theory solvers using this algorithm and describe the implementation and theory of this algorithm in detail. To extend the class of problems that the theory solver can robustly support, we borrow and adapt several techniques from linear programming (LP) and mixed integer programming (MIP) solvers which come from the tradition of optimization. We propose a new decicion procedure for quantifier-free linear real arithmetic that replaces the Simplex for DPLL(T) algorithm with a variant of the Simplex algorithm that performs a form of optimization ‒ minimizing the sum of infeasibilties. In this thesis, we additionally describe techniques for leveraging LP and MIP solvers to improve the performance of SMT solvers without compromising correctness. Previous efforts to leverage such solvers in the context of SMT have concluded that in addition to being potentially unsound, such solvers are too heavyweight to compete in the context of SMT. We present an empirical comparison against other state-of-the-art SMT tools to demonstrate the effectiveness of the proposed solutions.
Title: Local temporal reasoning
Author(s): Koskinen, Eric
Abstract:
We present the first method for reasoning about temporal logic properties of higher-order, infinite-data programs. By distinguishing between the finite traces and infinite traces in the specification, we obtain rules that permit us to reason about the temporal behavior of program parts via a type-and-effect system, which is then able to compose these facts together to prove the overall target property of the program. The type system alone is strong enough to derive many temporal safety properties using refinement types and temporal effects. We also show how existing techniques can be used as oracles to provide liveness information (e.g. termination) about program parts and that the type-and-effect system can combine this information with temporal safety information to derive nontrivial temporal properties. Our work has application toward verification of higher-order software, as well as modular strategies for procedural programs.
Title: The Push/Pull model of transactions
Author(s): Koskinen, Eric; Parkinson, Matthew
Abstract:
We present a general theory of serializability, unifying a wide range of transactional algorithms, including some that are yet to come. To this end, we provide a compact semantics in which concurrent transactions push their effects into the shared view (or unpush to recall effects) and pull the effects of potentially uncommitted concurrent transactions into their local view (or unpull to detangle). Each operation comes with simple side-conditions given in terms of commutativity (Lipton's left-movers and right-movers).
The benefit of this model is that most of the elaborate reasoning (coinduction, simulation, subtle invariants, etc.) necessary for proving the serializability of a transactional algorithm is already proved within the semantic model. Thus, proving serializability (or opacity) amounts simply to mapping the algorithm on to our rules, and showing that it satisfies the rules' side-conditions.
Title: Cryptographic Algorithms for the SecureDelegation of Multiparty Computation
Candidate: Lopez-Alt, Adriana
Advisor(s): Dodis, Yevgeniy
Abstract:
In today’s world, we store our data and perform expensive computations remotely on powerful servers (a.k.a. “the cloud”) rather than on our local devices. In this dissertation we study the question of achieving cryptographic security in the setting where multiple (mutually distrusting) clients wish to delegate the computation of a joint function on their inputs to an untrusted cloud, while keeping these inputs private. We introduce two frameworks for modeling such protocols.
We construct cloud-assisted and on-the-fly MPC protocols using fully homomorphic encryption (FHE). However, FHE requires inputs to be encrypted under the same key; we extend it to the multiparty setting in two ways:
Title: Resolution-Exact Planner for a 2-link Planar Robot using Soft Predicates
Candidate: Luo, Zhongdi
Advisor(s): Yap, Chee
Abstract:
Motion planning is a major topic in robotics. It frequently refers to motion of a robot in a R 2 or R 3 world that contains obstacles. Our goal is to produce algorithms that are practical and have strong theoretical guarantees. Recently, a new framework Soft Subdivision Search (SSS) was introduced to solve various motion planning problems. It is based on soft predicates and a new notion of correctness called resolution-exactness. Unlike most theoretical algorithms, such algorithms can be implemented without exact computation. In this thesis we describe a detailed, realized algorithm of SSS for a 2-link robot in R 2 . We prove the correctness of our predicates and also do experimental study of several strategies to enhance the basic SSS algorithm. In particular, we introduce a technique called T/R Splitting, in which the splittings of the rotational degrees of freedom are deferred to the end. The results give strong evidence of the practicability of SSS.
Title: Robust and Efficient Methods for Approximation and Optimization of Stability Measures
Candidate: Mitchell, Tim
Advisor(s): Overton, Michael
Abstract:
We consider two new algorithms with practical application to the problem of designing controllers for linear dynamical systems with input and output: a new spectral value set based algorithm called hybrid expansion-contraction intended for approximating the H-infinity norm, or equivalently, the complex stability radius, of large-scale systems, and a new BFGS SQP based optimization method for nonsmooth, nonconvex constrained optimization motivated by multi-objective controller design. In comprehensive numerical experiments, we show that both algorithms in their respect domains are significantly faster and more robust compared to other available alternatives. Moreover, we present convergence guarantees for hybrid expansion-contraction, proving that it converges at least superlinearly, and observe that it converges quadratically in practice, and typically to good approximations to the H-infinity norm, for problems which we can verify this. We also extend the hybrid expansion-contraction algorithm to the real stability radius, a measure which is known to be more difficult to compute than the complex stability radius. Finally, for the purposes of comparing multiple optimization methods, we present a new visualization tool called relative minimization profiles that allow for simultaneously assessing the relative performance of algorithms with respect to three important performance characteristics, highlighting how these measures interrelate to one another and compare to the other competing algorithms on heterogenous test sets. We employ relative minimization profiles to empirically validate our proposed BFGS SQP method in terms of quality of minimization, attaining feasibility, and speed of progress compared to other available methods on challenging test sets comprised of nonsmooth, nonconvex constrained optimization problems arising in controller design.
Title: Building Efficient Distributed In-memory Systems
Candidate: Power, Russell
Advisor(s): Li, Jinyang
Abstract:
The recent cloud computing revolution has changed the distributed computing landscape, making the resources of entire datacenters available to ordinary users. This process has been greatly aided by dataflow style frameworks such as MapReduce which expose simple model for programs, allowing for efficient, fault-tolerant execution across many machines. While the MapReduce model has proved to be effective for many applications, there are a wide class of applications which are difficult to write or inefficient in such a model. This includes many familiar and important applications such as PageRank, matrix factorization and a number of machine learning algorithms. In lieu of a good framework for building these applications, users resort to writing applications using MPI or RPC, a difficult and error-prone construction.
This thesis presents 2 complementary frameworks, Piccolo and Spartan, which help programmers to write in-memory distributed applications not served well by existing approaches.
Piccolo presents a new data-centric programming model for in-memory applications. Unlike data-flow models, Piccolo allows programs running on different machines to share distributed, mutable state via a key-value table interface. This design allows for both high-performance and additional flexibility. Piccolo makes novel use of commutative updates to efficiently resolve write-write conflicts. We find Piccolo provides an efficient backend for a wide-range of applications: from PageRank and matrix multiplication to web-crawling.
While Piccolo provides an efficient backend for distributed computation, it can still be some- what cumbersome to write programs using it directly. To address this, we created Spartan. Spartan implements a distributed implementation of the NumPy array language, and fully sup- ports important array language features such as spatial indexing (slicing), fancy indexing and broadcasting. A key feature of Spartan is its use of a small number of simple, powerful high-level operators to provide most functionality. Not only do these operators dramatically simplify the design and implementation of Spartan, they also allow users to implement new functionality with ease.
We evaluate Piccolo and Spartan on a wide range of applications and find that they both perform significantly better than existing approaches.
Title: VerifiableAuction: An Auction System for a Suspicious World
Author(s): Rosenberg, Michael; Shasha, Dennis
Abstract:
This paper presents a cryptosystem that will allow for fair first-price sealed-bid auctions among groups of individuals to be conducted over the internet without the need for a trusted third party. A client who maintains the secrecy of his or her private key will be able to keep his/her bid secret from the server and from all other clients until this client explicitly decides to reveal his/her bid, which will be after all clients publish their obfuscated bids. Each client will be able to verify that every other client's revealed bid corresponds to that client's obfuscated bid at the end of each auction. Each client is provided with a transcript of all auction proceedings so that they may be independently audited.
Title: Runtime Compilation of Array-Oriented Python Programs
Candidate: Rubinsteyn, Alex
Advisor(s): Shasha, Dennis
Abstract:
The Python programming language has become a popular platform for data analysis and scientific computing. To mitigate the poor performance of Python's standard interpreter, numerically intensive computations are typically offloaded to library functions written in languages such as Fortran or C. If, however, some algorithm does not have an existing low-level implementation, then the scientific programmer must either accept sub-standard performance (sometimes orders of magnitude slower than native code) or themselves implement the desired functionality in a less productive but more efficient language.
To alleviate this problem, this thesis present Parakeet, a runtime compiler for an array-oriented subset of Python. Parakeet does not replace the Python interpreter, but rather selectively augments it by compiling and executing functions explicitly marked by the programmer. Parakeet uses runtime type specialization to eliminate the performance-defeating dynamicism of untyped Python code. Parakeet's pervasive use of data parallel operators as a means for implementing array operations enables high-level restructuring optimization and compilation to parallel hardware such as multi-core CPUs and graphics processors. We evaluate Parakeet on a collection of numerical benchmarks and demonstrate its dramatic capacity for accelerating array-oriented Python programs.
Title: A Deep Learning Pipeline for Image Understanding and Acoustic Modeling
Candidate: Sermanet, Pierre
Advisor(s): LeCun, Yann
Abstract:
One of the biggest challenges artificial intelligence faces is making sense of the real world through sensory signals such as audio or video. Noisy inputs, varying object viewpoints, deformations and lighting conditions turn it into a high-dimensional problem which cannot be efficiently solved without learning from data.
This thesis explores a general way of learning from high dimensional data (video, images, audio, text, financial data, etc.) called deep learning. It strives on the increasingly large amounts of data available to learn robust and invariant internal features in a hierarchical manner directly from the raw signals.
We propose an unified pipeline for feature learning, recognition, localization and detection using Convolutional Networks (ConvNets) that can obtain state-of-the-art accuracy on a number of pattern recognition tasks, including acoustic modeling for speech recognition and object recognition in computer vision. ConvNets are particularly well suited for learning from continuous signals in terms of both accuracy and efficiency.
Additionally, a novel and general deep learning approach to detection is proposed and successfully demonstrated on the most challenging vision datasets. We then generalize it to other modalities such as speech data. This approach allows accurate localization and detection objects in images or phones in voice signals by learning to predict boundaries from internal representations. We extend the reach of deep learning from classification to detection tasks in an integrated fashion by learning multiple tasks using a single deep model. This work is among the first to outperform human vision and establishes a new state of the art on some computer vision and speech recognition benchmarks.
Title: Towards New Interfaces For Pedagogy
Candidate: Stein, Murphy
Advisor(s): Perlin, Ken
Abstract:
Developing technology to help people teach and learn is an important topic in Human Computer Interaction (HCI).
In this thesis we present three studies on this topic. In the first study, we demonstrate new games for learning mathematics and discuss the evidence for key design decisions from user studies. In the second study, we develop a real-time video compositing system for distance education and share evidence for its potential value compared to standard techniques from two user studies. In the third study, we demonstrate our markerless hand tracking interface for real-time 3D manipulation and explain its advantages compared to other state-of-the-art methods.
A data-driven methodology is applied intensively throughout the course of this study. Several paraphrase corpora are constructed using automatic techniques, experts and crowdsourcing platforms. Paraphrase systems are trained and evaluated by using these data as a cornerstone. We show that even with a very noisy or a relatively small amount of parallel training data, it is possible to learn paraphrase models which capture linguistic phenomena. This work expands the scope of paraphrase studies to targeting different language variations, and more potential applications, such as text normalization and domain adaptation.
Title: Computational Complexity Implicationsof Secure Coin-Flipping
Candidate: Tentes, Aristeidis
Advisor(s): Dodis, Yevgeniy
Abstract:
Modern Cryptography is based on computational intractability assumptions, e.g., Factoring, Discrete Logarithm, Diffie-Helman etc. However, since an assumption might be proven incorrect, there has been a lot of focus in order to construct cryptographic primitives based on the possibly most minimal assumption. The most popular minimal assumption, which is implied by the existence of almost all cryptographic primitives, is the existence of One Way Functions. Coin-Flipping protocols are known to be implied by One-Way Functions, however, a complete characterization of the inverse direction is not known. There was even speculation that weak notions of Coin Flipping Protocols might be strictly weaker than One Way Functions. In this thesis we show that even very weak notions of Coin Flipping protocols do imply One Way Functions. In particular we show that the existence of a coin-flipping protocol safe against any non-trivial constant bias (e.g 0.499) implies the existence of One Way Functions. This improves upon a recent result of Haitner and Omri [FOCS '11], who proved this implication for protocols with bias 0.207. Unlike the former result, our result also holds for weak coin-flipping protocols.
Title: On Automating Separation Logic with Trees and Data
Author(s): Wies, Thomas
Abstract:
Separation logic (SL) is a widely used formalism for verifying heap manipulating programs. Existing SL solvers focus on decidable fragments for list-like structures. More complex data structures such as trees are typically unsupported in implementations, or handled by incomplete heuristics.
While complete decision procedures for reasoning about trees have been proposed, these procedures suffer from high complexity, or make global assumptions about the heap that contradict the separation logic philosophy of local reasoning. In this paper, we present a fragment of classical first-order logic for local reasoning about tree-like data structures. The logic is decidable in NP and the decision procedure allows for combinations with other decidable first-order theories for reasoning about data. Such extensions are essential for proving functional correctness properties.
We have implemented our decision procedure and, building on earlier work on translating SL proof obligations into classical logic, integrated it into an SL-based verification tool. We successfully used the tool to verify functional correctness of tree-based data structure implementations.
Title: Data-driven Approaches for Paraphrasing across Language Variations
Candidate: Xu, Wei
Advisor(s): Grishman, Ralph
Abstract:
Our language changes very rapidly, accompanying political, social and cultural trends, as well as the evolution of science and technology. The Internet, especially the social media, has accelerated this process of change. This poses a severe challenge for both human beings and natural language processing (NLP) systems, which usually only model a snapshot of language presented in the form of text corpora within a certain domain and time frame.
While much previous effort has investigated monolingual paraphrase and bilingual translation, we focus on modeling meaning-preserving transformations between variants of a single language. We use Shakespearean and Internet language as examples to investigate various aspects of this new paraphrase problem, including acquisition, generation, detection and evaluation.
A data-driven methodology is applied intensively throughout the course of this study. Several paraphrase corpora are constructed using automatic techniques, experts and crowdsourcing platforms. Paraphrase systems are trained and evaluated by using these data as a cornerstone. We show that even with a very noisy or a relatively small amount of parallel training data, it is possible to learn paraphrase models which capture linguistic phenomena. This work expands the scope of paraphrase studies to targeting different language variations, and more potential applications, such as text normalization and domain adaptation.
Title: Positive-Unlabeled Learning in the Context of Protein Function Prediction
Candidate: Youngs, Noah
Advisor(s): Shasha, Dennis
Abstract:
With the recent proliferation of large, unlabeled data sets, a particular subclass of semisupervised learning problems has become more prevalent. Known as positiveunlabeled learning (PU learning), this scenario provides only positive labeled examples, usually just a small fraction of the entire dataset, with the remaining examples unknown and thus potentially belonging to either the positive or negative class. Since the vast majority of traditional machine learning classifiers require both positive and negative examples in the training set, a new class of algorithms has been developed to deal with PU learning problems.
A canonical example of this scenario is topic labeling of a large corpus of documents. Once the size of a corpus reaches into the thousands, it becomes largely infeasible to have a curator read even a sizable fraction of the documents, and annotate them with topics. In addition, the entire set of topics may not be known, or may change over time, making it impossible for a curator to annotate which documents are NOT about certain topics. Thus a machine learning algorithm needs to be able to learn from a small set of positive examples, without knowledge of the negative class, and knowing that the unlabeled training examples may contain an arbitrary number of additional but as yet unknown positive examples. Another example of a PU learning scenario recently garnering attention is the protein function prediction problem (PFP problem).
While the number of organisms with fully sequenced genomes continues to grow, the progress of annotating those sequences with the biological functions that they perform lags far behind. Machine learning methods have already been successfully applied to this problem, but with many organisms having a small number of positive annotated training examples, and the lack of availability of almost any labeled negative examples, PU learning algorithms can make large gains in predictive performance.
The first part of this dissertation motivates the protein function prediction problem, explores previous work, and introduces novel methods that improve upon previously reported benchmarks for a particular type of learning algorithm, known as Gaussian Random Field Label Propagation (GRFLP). In addition, we present improvements to the computational efficiency of the GRFLP algorithm, and a modification to the traditional structure of the PFP learning problem that allows for simultaneous prediction across multiple species.
The second part of the dissertation focuses specifically on the positive-unlabeled aspects of the PFP problem. Two novel algorithms are presented, and rigorously compared to existing PU learning techniques in the context of protein function prediction. Additionally, we take a step back and examine some of the theoretical considerations of the PU scenario in general, and provide an additional novel algorithm applicable in any PU context. This algorithm is tailored for situations in which the labeled positive examples are a small fraction of the set of true positive examples, and where the labeling process may be subject to some type of bias rather than being a random selection of true positives (arguably some of the most difficult PU learning scenarios).
The third and fourth sections return to the PFP problem, examining the power of tertiary structure as a predictor of protein function, as well as presenting two case studies of function prediction performance on novel benchmarks. Lastly, we conclude with several promising avenues of future research into both PU learning in general, and the protein function prediction problem specifically.
Title: Hierarchical Convolutional Deep Learning in Computer Vision
Candidate: Zeiler, Matthew
Advisor(s): Fergus, Rob
Abstract:
It has long been the goal in computer vision to learn a hierarchy of features useful for object recognition. Spanning the two traditional paradigms of machine learning, unsupervised and supervised learning, we investigate the application of deep learning methods to tackle this challenging task and to learn robust representations of images.
We begin our investigation with the introduction of a novel unsupervised learning technique called deconvolutional networks. Based on convolutional sparse coding, we show this model learns interesting decompositions of images into parts without object label information. This method, which easily scales to large images, becomes increasingly invariant by learning multiple layers of feature extraction coupled with pooling layers. We introduce a novel pooling method called Gaussian pooling to enable these layers to store continuous location information while being differentiable, creating a unified objective function to optimize.
In the supervised learning domain, a well-established model for recognition of objects is the convolutional network. We introduce a new regularization method for convolutional networks called stochastic pooling which relies on sampling noise to prevent these powerful models from overfitting. Additionally, we show novel visualizations of these complex models to better understand what they learn and to provide insight on how to develop state-of-the-art architectures for large-scale classification of 1,000 different object categories.
We also investigate some other related problems in deep learning. First, we introduce a model for the task of mapping one high dimensional time series sequence onto another. Second, we address the choice of nonlinearity in neural networks, showing evidence that rectified linear units outperform others types in automatic speech recognition. Finally, we introduce a novel optimization method called ADADELTA which shows promising convergence speeds in practice while being robust to hyper-parameter selection.
Title: Isogeometric BDDC Preconditioners with Deluxe Scaling
Author(s): Beirao Da Veiga, Lourenco; Pavarino, Luca; Scacchi, Simone; Widlund, Olof; Zampni, Stefano
Abstract:
A BDDC (Balancing Domain Decomposition by Constraints) preconditioner with a novel scaling, introduced by Dohrmann for problems with more than one variable coeffcient and here denoted as deluxe scaling, is extended to Isogeometric Analysis of scalar elliptic problems. This new scaling turns out to be more powerful than the standard rho- and stiffness scalings considered in a previous isogeometric BDDC study. Our h-analysis shows that the condition number of the resulting deluxe BDDC preconditioner is scalable with a quasi-optimal polylogarithmic bound which is also independent of coeffcient discontinuities across subdomain interfaces. Extensive numerical experiments support the theory and show that the deluxe scaling yields a remarkable improvement over the older scalings, in particular, for large isogeometric polynomial degree and high regularity.
Title: An Efficient Active Learning Framework for New Relation Types
Candidate: Fu, Lisheng
Advisor(s): Grishman, Ralph; Davis, Ernest
Abstract:
Relation extraction is a fundamental task in information extraction. Different methods have been studied for building a relation extraction system. Supervised training of models for this task has yielded good performance, but at substantial cost for the annotation of large training corpora (About 40K same-sentence entity pairs). Semi-supervised methods can only require a seed set, but the performance is very limited when the seed set is very small, which is not very satisfactory for real relation extraction applications. The trade-off of annotation and performance is also hard to decide in practice. Active learning strategies allow users to gradually improve the model and to achieve comparable performance to supervised methods with limited annotation. Recent study shows active learning on this task needs much fewer labels for each type to build a useful relation extraction application. We feel active learning is a good direction to do relation extraction and presents a more efficient active learning framework. This framework starts from a better balance between positive and negative samples, and boosts by interleaving self-training and co-testing. We also studied the reduction of annotation cost by enforcing argument type constraints. Experiments show a substantial speed-up by comparison to previous state-of-the-art pure co-testing active learning framework. We obtain reasonable performance with only a hundred labels for individual ACE 2004 relation types. We also developed a GUI tool for real human-in-the-loop active learning trials. The goal of building relation extraction systems in a very short time seems to be promising.
Title: Incentive-Centered Design of Money-Free Mechanisms
Candidate: Gkatzelis, Vasilis
Advisor(s): Cole, Richard
Abstract:
This thesis serves as a step toward a better understanding of how to design fair and efficient multiagent resource allocation systems by bringing the incentives of the participating agents to the center of the design process. As the quality of these systems critically depends on the ways in which the participants interact with each other and with the system, an ill-designed set of incentives can lead to severe inefficiencies. The special focus of this work is on the problems that arise when the use of monetary exchanges between the system and the participants is prohibited. This is a common restriction that substantially complicates the designer's task; we nevertheless provide a sequence of positive results in the form of mechanisms that maximize efficiency or fairness despite the possibly self-interested behavior of the participating agents.
The first part of this work is a contribution to the literature on approximate mechanism design without money. Given a set of divisible resources, our goal is to design a mechanism that allocates them among the agents. The main complication here is due to the fact that the agents' preferences over different allocations of these resources may not be known to the system. Therefore, the mechanism needs to be designed in such a way that it is in the best interest of every agent to report the truth about her preferences; since monetary rewards and penalties cannot be used in order to elicit the truth, a much more delicate regulation of the resource allocation is necessary. Our contribution mostly revolves around a new truthful mechanism that we propose, which we call the /Partial Allocation/ mechanism. We first show how to use the two-agent version of this mechanism to create a system with the best currently known worst-case efficiency guarantees for problem instances involving two agents. We then consider fairness measures and prove that the general version of this elegant mechanism yields surprisingly good approximation guarantees for the classic problem of fair division. More specifically, we use the well established solution of /Proportional Fairness/ as a benchmark and we show that for an arbitrary number of agents and resources, and for a very large class of agent preferences, our mechanism provides /every agent/ with a value close to her proportionally fair value. We complement these results by also studying the limits of truthful money-free mechanisms, and by providing other mechanisms for special classes of problem instances. Finally, we uncover interesting connections between our mechanism and the Vickrey-Clarke-Groves mechanism from the literature on mechanism design with money.
The second part of this work concerns the design of money-free resource allocation mechanisms for /decentralized/ multiagent systems. As the world has become increasingly interconnected, such systems are using more and more resources that are geographically dispersed; in order to provide scalability in these systems, the mechanisms need to be decentralized. That is, the allocation decisions for any given resource should not assume global information regarding the system's resources or participants. We approach this restriction by using /coordination mechanisms/: a collection of simple resource allocation policies, each of which controls only one of the resources and uses only local information regarding the state of the system. The system's participants, facing these policies, have the option of choosing which resources they will access. We study a variety of coordination mechanisms and we prove that the social welfare of any equilibrium of the games that these mechanisms induce is a good approximation of the optimal welfare. Once again, we complement our positive results by studying the limits of coordination mechanisms. We also provide a detailed explanation of the seemingly counter-intuitive incentives that some of these mechanisms yield. Finally, we use this understanding in order to design a combinatorial constant-factor approximation algorithm for maximizing the social welfare, thus providing evidence that a game-theoretic mindset can lead to novel optimization algorithms.
Title: Tight Lower Bound on the Probability of a Binomial Exceeding its Expectation
Author(s): Greenberg, Spencer; Mohri, Mehryar
Abstract:
We give the proof of a tight lower bound on the probability that a binomial random variable exceeds its expected value. The inequality plays an important role in a variety of contexts, including the analysis of relative deviation bounds in learning theory and generalization bounds for unbounded loss functions.
Title: Locality Optimization for Data Parallel Programs
Candidate: Hielscher, Eric
Advisor(s): Shasha, Dennis
Abstract:
Productivity languages such as NumPy and Matlab make it much easier to implement data-intensive numerical algorithms than it is to implement them in efficiency languages such as C++. This is important as many programmers (1) aren't expert programmers; or (2) don't have time to tune their software for performance, as their main job focus is not programming per se. The tradeoff is typically one of execution time versus programming time, as unless there are specialized library functions or precompiled primitives for your particular task a productivity language is likely to be orders of magnitude slower than an efficiency language.
In this thesis, we present Parakeet, an array-oriented language embedded within Python, a widely-used productivity language. The Parakeet just-in-time compiler dynamically translates whole user functions to high performance multi-threaded native code. This thesis focuses in particular on our use of data parallel operators as a basis for locality enhancing program optimizations. e transform Parakeet programs written with the classic data parallel operators (Map, Reduce, and Scan; in Parakeet these are called adverbs) to process small local pieces (called tiles) of data at a time. To express this locality we introduce three new adverbs: TiledMap, TiledReduce, and TiledScan. These tiled adverbs are not exposed to the programmer but rather are automatically generated by a tiling transformation.
We use this tiling algorithm to bring two classic locality optimizations to a data parallel setting: cache tiling, and register tiling. We set register tile sizes statically at compile time, but use an online autotuning search to find good cache tile sizes at runtime. We evaluate Parakeet and these optimizations on various benchmark programs, and exhibit excellent performance even compared to typical C implementations.
Title: Diet Planner: Finding a Nutritionally Sound Diet While Following (Most) of a Dieter’s Desires
Author(s): Jermsurawong, Mick Jermsak; Shasha, Dennis
Abstract:
We describe the design and implementation of a diet website currently housed at http://nutrientdata.herokuapp.com/. The site allows users or dieticians to enter nutritional constraints (e.g. at least this much calcium but not more than that amount of calcium) and objectives (e.g. minimize calories), a list of foods/brands the person likes. The site then determines, if possible, the quantities of at least some of those desired foods that would meet the nutritional constraints. If not possible, then the site guides the user in the choice of other foods that may meet the nutritional constraints. The net result is a tailored diet measured in servings.
Title: A General Method for Energy-Error Tradeoffs in Approximate Adders
Author(s): Kedem, Zvi; Muntimadugu, Kirthi Krishna
Abstract:
Approximate adders are adders with conventional architectures run in an overclocked mode. With this mode, erroneous sums may be produced at the savings of energy required to execute the computation. The results presented in this report lead to a procedure for allocating the available energy budgets among the adders modules so as to minimize the expected error. For simplicity, only the uniform distribution of the inputs is considered.
Title: Reversibility of Turing Machine Computations
Author(s): Kedem, Zvi M.
Abstract:
Since Bennett's 1973 seminal paper, there has been a growing interest in general-purpose, reversible computations and they have been studied using both mathematical and physical models. Following Bennett, given a terminating computation of a deterministic Turing Machine, one may be interested in constructing a new Turing Machine, whose computation consists of two stages. The first stage emulates the original Turing Machine computation on its working tape, while also producing the trace of the computation on a new history tape. The second stage reverses the first stage using the trace information. Ideally, one would want the second stage to traverse whole-machine states in the reverse order from that traversed in the first stage. But this is impossible other than for trivial computations. Bennett constructs the second phase by using additional controller states, beyond those used during the first stage. In this report, a construction of the new machine is presented in which the second stage uses the same and only those controller states that the first stage used and they are traversed in the reverse order. The sole element that is not fully reversed is the position of the head on the history tape, where it is out of phase by one square compared to the first stage.
Title: Piecewise Smooth Surfaces with Features
Candidate: Kovacs, Denis
Advisor(s): Zorin, Denis
Abstract:
The creation, manipulation and display of piecewise smooth surfaces has been a fundamental topic in computer graphics since its inception. The applications range from highest-quality surfaces for manufacturing in CAD, to believable animations of virtual creatures in Special Effects, to virtual worlds rendered in real-time in computer games.
Our focus is on improving the a) mathematical representation and b) automatic construction of such surfaces from finely sampled meshes in the presence of features. Features can be areas of higher geometric detail in an otherwise smooth area of the mesh, or sharp creases that contrast the overall smooth appearance of an object.
In the first part, we build on techniques that define piecewise smooth surfaces, to improve their quality in the presence of features. We present a crease technique suitable for real-time applications that helps increases the perceived visual detail of objects that are required to be very compactly represented and efficiently evaluated.
We then introduce a new subdivision scheme that allows the use of T-junctions for better local refinement. It thus reduces the need for extraordinary vertices, which can cause surface artifacts especially on animated objects.
In the second part, we consider the problem of how to build the control meshes of piecewise smooth surfaces, in a way that the resulting surface closely approximates an existing data set (such as a 3D range scan), particularly in the presence of features. To this end, we introduce a simple modification that can be applied to a wide range of parameterization techniques to obtain an anisotropic parameterization. We show that a resulting quadrangulation can indeed better approximate the original surface. Finally, we present a quadrangulation scheme that turns a data set into a quad mesh with T-junctions, which we then use as a T-Spline control mesh to obtain a smooth surface.
Title: Low-level Image Priors and Laplacian Preconditioners for Applications in Computer Graphics and Computational Photography
Candidate: Krishnan, Dilip
Advisor(s): Fergus, Rob
Abstract:
In the first part of this thesis, we develop novel image priors and efficient algorithms for image denoising and deconvolution applications. Our priors and algorithms enable fast, high-quality restoration of images corrupted by noise or blur. In the second part, we develop effective preconditioners for Laplacian matrices. Such matrices arise in a number of computer graphics and computational photography problems such as image colorization, tone mapping and geodesic distance computation on 3D meshes.
The first prior we develop is a spectral prior which models correlations between different spectral bands. We introduce a prototype camera and flash system, used in conjunction with the spectral prior, to enable taking photographs at very low light levels. Our second prior is a sparsity-based measure for blind image deconvolution. This prior gives lower costs to sharp images than blurred ones, enabling the use simple and efficient Maximum a-Posteriori algorithms.
We develop a new algorithm for the non-blind deconvolution problem. This enables extremely fast deconvolution of images blurred by a known blur kernel. Our algorithm uses Fast Fourier Transforms and Lookup Tables to achieve real-time deconvolution performance with non convex gradient-based priors. Finally, for certain image restoration problems with no clear formation model, we demonstrate how learning a direct mapping between original/corrupted patch pairs enables effective restoration.
We develop multi-level preconditioners to solve discrete Poisson equations. Existing multilevel preconditioners have two major drawbacks: excessive bandwidth growth at coarse levels; and the inability to adapt to problems with highly varying coefficients. Our approach tackles both these problems by introducing sparsification and compensation steps at each level. We interleave the selection of fine and coarse-level variables with the removal of weak connections between potential fine-level variables (sparsification) and compensate for these changes by strengthening nearby connections. By applying these operations before each elimination step and repeating the procedure recursively on the resulting smaller systems, we obtain highly efficient schemes. The construction is linear in time and memory. Numerical experiments demonstrate that our new schemes outperform state of the art methods, both in terms of operation count and wall-clock time, over a range of 2D and 3D problems.
Title: A Balancing Domain Decomposition By Constraints Deluxe Method For Numerically Thin Reissner-Mindlin Plates Approximated With Falk-tu Finite Elements
Author(s): Lee, Jong Ho
Abstract:
The Reissner-Mindlin plate models thin plates. The condition numbers of finite element approximations of these plate models increase very rapidly as the thickness of the plate goes to 0. A Balancing Domain Decomposition by Constraints (BDDC) De Luxe method is developed for these plate problems discretized by Falk-Tu finite elements. In this new algorithm, subdomain Schur complements restricted to individual edges are used to define the average operator for the BDDC De Luxe method. It is established that the condition number of this preconditioned iterative method is bounded by C(1 + log (H/h))^2 if t, the thickness of the plate, is on the order of the element size h or smaller; H is the maximum diameter of the subdomains. The constant C is independent of the thickness t as well as H and h. Numerical results, which verify the theory, and a comparison with a traditional BDDC method are also provided.
Title: Cryptographic Security of Macaroon Authorization Credentials
Author(s): Lopez-Alt, Adriana
Abstract:
Macaroons, recently introduced by Birgisson et al., are authorization credentials that provide support for controlled sharing in decentralized systems. Macaroons are similar to cookies in that they are bearer credentials, but unlike cookies, macaroons include caveats that attenuate and contextually confine when, where, by who, and for what purpose authorization should be granted.
In this work, we formally study the cryptographic security of macaroons. We define macaroon schemes, introduce corresponding security definitions and provide several constructions. In particular, the MAC-based and certificate-based constructions outlined by Birgisson et al., can be seen as instantiations of our definitions. We also present a new construction that is privately-verifiable (similar to the MAC-based construction) but where the verifying party does not learn the intermediate keys of the macaroon, a problem already observed by Birgisson et al.
We also formalize the notion of a protocol for "discharging" third-party caveats and present a security definition for such a protocol. The encryption-based protocol outlined by Birgisson et al. can be seen as an instantiation of our definition, and we also present a new signature-based construction.
Finally, we formally prove the security of all constructions in the given security models.
Title: Relation Extraction with Weak Supervision and Distributional Semantics
Candidate: Min, Bonan
Advisor(s): Grishman, Ralph
Abstract:
Relation Extraction aims at detecting and categorizing semantic relations between pairs of entities in unstructured text. It benefits an enormous number of applications such as Web search and Question Answering. Traditional approaches for relation extraction either rely on learning from a large number of accurate human-labeled examples or pattern matching with hand-crafted rules. These resources are very laborious to obtain and can only be applied to a narrow set of target types of interest.
This talk focuses on learning relations with little or no human supervision. First, we examine the approach that treats relation extraction as a supervised learning problem. We develop an algorithm that is able to train a model with approximately 1/3 of the human-annotation cost and that matches the performance of models trained with high-quality annotation. Second, we investigate distant supervision, a weakly supervised algorithm that automatically generates its own labeled training data. We develop a latent Bayesian framework for this purpose. By using a model which provides a better approximation of the weak source of supervision, it outperforms the state-of-the-art methods. Finally, we investigate the possibility of building all relational tables beforehand with an unsupervised relation extraction algorithm. We develop an effective yet efficient algorithm that combines the power of various semantic resources that are automatically mined from a corpus based on distributional semantics. The algorithm is able to extract a very large set of relations from the web at high precision.
Title: Foundations of a Formal Theory of Time Travel
Author(s): Morgenstern, Leora
Abstract:
Although the phenomenon of time travel is common in popular culture, there has been little work in AI on developing a formal theory of time travel. This paper develops such a theory. The paper introduces a branching-time ontology that maintains the classical restriction of forward movement through a temporal tree structure, but permits the representation of paths in which one can perform inferences about time-travel scenarios. Central to the ontology is the notion of an agent embodiment whose beliefs are equivalent to those of an agent who has time-traveled from the future. We show how to formalize an example scenario and demonstrate what it means for such a scenario to be motivated with respect to an agent embodiment.
Title: A BDDC Algorithm for Raviart-Thomas Vector Fields
Author(s): Oh, Duk-Soon; Widlund, Olof B.; Dohrmann, Clark R.
Abstract:
A BDDC preconditioner is defined by a coarse component, expressed in terms of primal constraints and a weighted average across the interface between the subdomains, and local components given in terms of Schur complements of local subdomain problems. A BDDC method for vector field problems discretized with Raviart-Thomas finite elements is introduced. Our method is based on a new type of weighted average developed to deal with more than one variable coefficient. A bound on the condition number of the preconditioned linear system is also provided which is independent of the values and jumps of the coefficients across the interface and has a polylogarithmic condition number bound in terms of the number of degrees of freedom of the individual subdomains. Numerical experiments for two and three dimensional problems are also presented, which support the theory and show the effectiveness of our algorithm even for certain problems not covered by our theory.
Title: Usable Security Mechanisms in the Developing World
Candidate: Paik, Michael
Advisor(s): Subramanian; Lakshminarayanan
Abstract:
Security and privacy are increasingly important in our interconnected world. Cybercrimes, including identity theft, phishing, and other attacks, are on the rise, and computer-assisted crimes such as theft and stalking are becoming commonplace.
Contemporary with this trend is the uptake of technology in the developing world, proceeding at a pace often outstripping that of the developed world. Penetration of mobile phones and services such as healthcare delivery, mobile money, and social networking is higher than that of even amenities like electricity. Connectivity is empowering disenfranchised people, providing information and services to the heretofore disconnected poor.
There are efforts to use technology to enhance physical security and well-being in the developing world, including citizen journalism, education, improving drug security, attendance tracking, etc.
However, there are significant challenges to security both in the digital and the physical domains that are particular to these contexts. Infrastructure is constrained, literacy, numeracy, and familiarity with basic technologies cannot be assumed, and environments are harsh on hardware. These circumstances often prevent security best practices from being transplanted directly to these regions â in many ways, the adoption of technology has overtaken the users ability to use it safely, and their trust in it is oftentimes reater than it should be.
This dissertation describes several systems and methodologies designed to operate in the developing world, using technologies and metaphors that are familiar to users and that are robust against the operating environments.
It begins with an overview of the state of affairs, and several threat models. It continues with a description of Signet, a method to use SIM cards as trusted computing hardware to provide secure signed receipts. Next, Epothecary describes a low-infrastructure system for tracking pharmaceuticals that also significantly and asymmetrically increases costs for counterfeiters. The balance consists of a description of a low-cost Biometric Terminal currently in use by NGOs in India performing DOTS-based tuberculosis treatment, Blacknoise, an investigation into the use of low-cost cameraphones with noisy imaging sensors for image-based steganography, and finally Innoculous, a low-cost, crowdsourcing system for combating the spread of computer viruses, particularly among non-networked computers, while also collecting valuable "epidemiological" data.
Title: Automating Separation Logic Using SMT
Author(s): Piskac, Ruzica; Wies, Thomas; Zufferey, Damien
Abstract:
Separation logic (SL) has gained widespread popularity because of its ability to succinctly express complex invariants of a program's heap configurations. Several specialized provers have been developed for decidable SL fragments. However, these provers cannot be easily extended or combined with solvers for other theories that are important in program verification, e.g., linear arithmetic. In this paper, we present a reduction of decidable SL fragments to a decidable first-order theory that fits well into the satisﬁability modulo theories (SMT) framework. We show how to use this reduction to automate satisfiability, entailment, frame inference, and abduction problems for separafion logic using SMT solvers. Our approach provides a simple method of integrating separation logic into existing verification tools that provide SMT backends, and an elegant way of combining SL fragments with other decidable first-order theories. We implemented this approach in a verification tool and applied it to heap-manipulating programs whose verification involves reasoning in theory combinations.
Title: Inapproximability Reductions and Integrality Gaps
Candidate: Popat, Preyas
Advisor(s): Khot, Subhash
Abstract:
In this thesis we prove intractability results for several well studied problems in combinatorial optimization.
Closest Vector Problem with Preprocessing (CVPP): We show that the preprocessing version of the well known Closest Vector Problem is hard to approximate to an almost polynomial factor unless NP is in quasi polynomial time. The approximability of CVPP is closely related to the security of lattice based cryptosystems.
Pricing Loss Leaders: We show hardness of approximation results for the problem of maximizing profit from buyers with single minded valuations where each buyer is interested in bundles of at most k items, and the items are allowed to have negative prices ("Loss Leaders"). For k = 2, we show that assuming the Unique Games Conjecture, it is hard to approximate the profit to any constant factor. For k > 2, we show the same result assuming P != N P.
Integrality gaps: We show SemiDefinite Programming (SDP) integrality gaps for Unique Games and 2 to 1 Games. Inapproximability results for these problems imply inapproximability results for many fundamental optimization problems. For the first problem, we show "approximate" integrality gaps for super constant rounds of the powerful Lasserre hierarchy. For the second problem we show integrality gaps for the basic SDP relaxation with perfect completeness.
Title: Natural Interaction with a Virtual World
Candidate: Rosenberg, Ilya
Advisor(s): Perlin, Ken
Abstract:
A large portion of computer graphics and human/computer interaction is concerned with the creation, manipulation and use of two and three dimensional objects existing in a virtual world. By creating more natural physical interfaces and virtual worlds which behave in physically plausible ways, it is possible to empower nonexpert users to create, work and play in virtual environments. This thesis is concerned with the design, creation, and optimization of user-input devices which break down the barriers between the real and the virtual as well as the development of software algorithms which allow for the creation of physically realistic virtual worlds.
Title: Parsing and Analyzing POSIX API behavior on different platforms
Candidate: Savvides, Savvas
Advisor(s): Cappos, Justin; Li, Jinyang
Abstract:
Because of the increased variety of operating systems and architectures, developing applications that are supported by multiple platforms has become a cumbersome task. To mitigate this problem, many portable APIs were created which are responsible for hiding the details of the underlying layers of the system and they provide a universal interface on top of which applications can be built. Many times it is necessary to examine the interactions between an application and an API, either to check that the behavior of these interactions is the expected one or to confirm that this behavior is the same across platforms. In this thesis, the POSIX Omni Tracer tool is described. POSIX Omni Tracer provides an easy way to analyze the behavior of the POSIX API under various platforms. The behavior of the POSIX API can be captured in traces during an application�۪s execution using various tracing utilities. These traces are then parsed into a uniform representation. Since the captured behavior from different platforms share the same format, they can be easily analyzed or compared between them.
Title: Security Mechanisms for Physical Authentication
Candidate: Sharma, Ashlesh
Advisor(s): Subramanian; Lakshminarayanan
Abstract:
Counterfeiting of goods is a worldwide problem where the losses are in billions of dollars. It is estimated that 10% of all the world trade is counterfeit. To alleviate counterfeiting, a number of techniques are used from barcodes to holograms. But these technologies are easily reproducible and hence they are ineffective against counterfeiters.
In this thesis, we introduce PaperSpeckle, a novel way to fingerprint any piece of paper based on its unique microscopic properties. Next, we extend and generalize this work to introduce TextureSpeckle, a novel way to fingerprint and characterize the uniqueness of the surface of a material based on the interaction of light with the natural randomness present in the rough structure at the microscopic level of the surface. We show the existence and uniqueness of these fingerprints by analyzing a large number of surfaces (over 20,000 microscopic surfaces and 200 million pairwise comparisons) of different materials. We also define the entropy of the fingerprints and show how each surface can be uniquely identified in a robust manner even in case of damage.
From a theoretical perspective, we consider a discrete approximation model from light scattering theory which allows us to compute the speckle pattern for a given surface. Under this computational model, we show that given a speckle pattern, it is computationally hard to reconstruct the physical surface characteristics by simulating the multiple scattering of light. Using TextureSpeckle as a security primitive, we design secure protocols to enable a variety of scenarios such as: i) supply chain security, where applications range from drug tracking to inventory management, ii) mobile based secure transfer of money (mobile money), where any paper can be changed to an on-demand currency, and iii) fingerprint ecosystem, a cloud based system, where any physical object can be identified and authenticated on-demand.
We discuss the construction of the prototype device ranging from optical lens design to usability aspects and show how our technique can be applied in the real world to alleviate counterfeiting and forgery. In addition, we introduce Pattern Matching Puzzles (PMPs), a usable security mechanism that provides a 'human computable' one-time-MAC (message authentication code) for every transaction,making each transaction information-theoretically secure against various adversarial attacks. The puzzles are easy tosolve even for semi-literate users with simple pattern recognition skills.
Title: Online Machine Learning Algorithms For Currency Exchange Prediction
Author(s): Soulas, Eleftherios; Shasha, Dennis
Abstract:
Using Machine Learning Algorithms to analyze and predict security price patterns is an area of active interest. Most practical stock traders combine computational tools with their intuitions and knowledge to make decisions.
This document explains the algorithms and discusses various metrics of accuracy. It validates the models by applying the model to a real-life trading price stream. Though it is very hard to replace the expertise of an experienced trader, software like this may enhance the trader's performance.
Title: Augmenting Information Flow for Visual Privacy
Candidate: Spiro, Ian
Advisor(s): Bregler, Christopher
Abstract:
In the Information Age, visual media take on powerful new forms. Photographs once printed on paper and stored in physical albums now exist as digital files. With the rise of social media, photo data has moved to the cloud for rapid dissemination. The upside can be measured in terms of increased efficiency, greater reach, or reduced printing costs. But there is a downside that is harder to quantify: the risk of private photos or videos leaking inappropriately. Human imagery is potentially sensitive, revealing private details of a persons body, lifestyle, activities, and more. Images create visceral responses and have the potential to permanently damage a persons reputation.
We employed the theory of contextual integrity to explore privacy aspects of transmitting the human form. In response to privacy threats from new sociotechnical systems, we developed practical solutions that have the potential to restore balance. The main work is a set of client-side, technical interventions that can be used to alter information flows and provide features to support visual privacy. In the first approach, we use crowdsourcing to extract specific, useful human signal from video to decouple it from bundled identity information. The second approach is an attempt to achieve similar ends with pure software. Instead of using information workers, we developed a series of filters that alter video to hide identity information while still revealing motion signal. The final approach is an attempt to control the recipients of photos by encoding them in the visual channel. The software completely protects data from third-parties who lack proper credentials and maintains data integrity by exploiting the visual coherence of uploaded images, even in the face of JPEG compression. The software offers end-to-end encryption that is compatible with existing social media applications.
Title: Toward a computational solution to the inverse problem of how hypoxia arises in metabolically heterogeneous cancer cell populations
Candidate: Sundstrom, Andrew
Advisor(s): Mishra, Bud; Bar-Sagi, Dafna
Abstract:
As a tumor grows, it rapidly outstrips its blood supply, leaving portions of tumor that undergo hypoxia. Hypoxia is strongly correlated with poor prognosis as it renders tumors less responsive to chemotherapy and radiotherapy. During hypoxia, HIFs upregulate production of glycolysis enzymes and VEGF, thereby promoting metabolic heterogeneity and angiogenesis, and proving to be directly instrumental in tumor progression. Prolonged hypoxia leads to necrosis, which in turn activates inflammatory responses that produce cytokines that stimulate tumor growth. Hypoxic tumor cells interact with macrophages and fibroblasts, both involved with inflammatory processes tied to tumor progression. So it is of clinical and theoretical significance to understand: Under what conditions does hypoxia arise in a heterogeneous cell population? Our aim is to transform this biological origins problem into a computational inverse problem, and then attack it using approaches from computer science. First, we develop a minimal, stochastic, spatiotemporal simulation of large heterogeneous cell populations interacting in three dimensions. The simulation can manifest stable localized regions of hypoxia. Second, we employ and develop a variety of algorithms to analyze histological images of hypoxia in xenographed colorectal tumors, and extract features to construct a spatiotemporal logical characterization of hypoxia. We also consider characterizing hypoxia by a linear regression functional learning mechanism that yields a similarity score. Third, we employ a Bayesian statistical model checking algorithm that can determine, over some bounded number of simulation executions, whether hypoxia is likely to emerge under some fixed set of simulation parameters, and some fixed logical or functional description of hypoxia. Driving the model checking process is one of three adaptive Monte Carlo sampling algorithms we developed to explore the high dimensional space of simulation initial conditions and operational parameters. Taken together, these three system components formulate a novel approach to the inverse problem above, and constitute a design for a tool that can be placed into the hands of experimentalists, for testing hypotheses based upon known parameter values or ones the tool might discover. In principle, this design can be generalized to other biological phenomena involving large heterogeneous populations of interacting cells.
Title: PhyloBrowser: A visual tool to explore phylogenetic trees
Candidate: Tershakovec, Tamara
Advisor(s): Shasha, Dennis; Coruzzi, Gloria
Abstract:
Primary acknowledgements go to my research advisor, Dennis Shasha, for his patient and unwavering support. I would also like to thank my second reader, Gloria Coruzzi, for encouraging and showcasing my work. Kranthi Varala helped immensely in explaining the biological and statistical concepts involved in this project. The Virtual Plant team got me started in biological data visualization and was a joy to work with. And finally, thanks to Chris Poultney, who put me in touch with Professor Shasha and started me on my way back to NYU.
Title: Rethinking Information Privacy for the Web
Candidate: Tierney, Matthew
Advisor(s): Subramanian; Lakshminarayanan
Abstract:
In response to Supreme Court Justice Samuel Alitoâs opinion that society should accept a decline in personal privacy with modern technology, Hanni M. Fakhoury, staff attorney with the Electronic Frontier Foundation, argued âTechnology doesnât involve an âinevitableâ tradeoff [of increased convenience] with privacy. The only inevitability must be the demand that privacy be a value built into our technologyâ [42]. Our position resonates with Mr. Fakhouryâs. In this thesis, we present three artifacts that address the balance between usability, efficiency, and privacy as we rethink information privacy for the web.
In the first part of this thesis, we present the design, implementation and evaluation of Cryptagram, a system designed to enhance online photo privacy. Cryptagram enables users to convert photos into encrypted images, which the users upload to Online Social Networks (OSNs). Users directly manage access control to those photos via shared keys that are independent of OSNs or other third parties. OSNs apply standard image transformations (JPEG compression) to all uploaded images so Cryptagram provides image encoding and encryption protocols that are tolerant to these transformations. Cryptagram guarantees that the recipient with the right credentials can completely retrieve the original image from the transformed version of the uploaded encrypted image while the OSN cannot infer the original image. Cryptagramâs browser extension integrates seamlessly with preexisting OSNs, including Facebook and Google+, and currently has over 400 active users.
In the second part of this thesis, we present the design and implementation of Lockbox, a system designed to provide end-to-end private file-sharing with the convenience of Google Drive or Dropbox. Lockbox uniquely combines two important design points: (1) a federated system for detecting and recovering from server equivocation and (2) a hybrid cryptosystem over delta encoded data to balance storage and bandwidth costs with efficiency for syncing end-user data. To facilitate appropriate use of public keys in the hybrid cryptosystem, we integrate a service that we call KeyNet, which is a web service designed to leverage existing authentication media (e.g., OAuth, verified email addresses) to improve the usability of public key cryptography.
In the third part of this thesis, we present the design of Compass, which realizes the philosophical privacy framework of contextual integrity (CI) as a full OSN design. CI), which we believe better captures users privacy expectations in OSNs. In Compass, three properties hold: (a) users are associated with roles in specific contexts; (b) every piece of information posted by a user is associated with a specific context; (c) norms defined on roles and attributes of posts in a context govern how information is shared across users within that context. Given the definition of a context and its corresponding norm set, we describe the design of a compiler that converts the human-readable norm definitions to generate appropriate information flow verification logic including: (a) a compact binary decision diagram for the norm set; and (b) access control code that evaluates how a new post to a context will flow. We have implemented a prototype that shows how the philosophical framework of contextual integrity can be realized in practice to achieve strong privacy guarantees with limited additional verification overhead.
Title: PAC-Learning for Energy-based Models
Candidate: Zhang, Xiang
Advisor(s): LeCun, Yann; Sontag, David
Abstract:
In this thesis we prove that probably approximately correct (PAC) learning is guaranteed for the framework of energy-based models. Starting from the very basic inequalities, we establish our theory based on the existence of metric between hypothesis, to which the energy function is Lipschitz continuous. The result of the theory provides a new scheme of regularization called central regularization, which puts the effect of deep learning and feature learning in a new perspective. Experiments of this scheme shows that it achieved both good generalization error and testing error.
Title: Transaction chains: achieving serializability with low latency in geo-distributed storage systems
Author(s): Zhang, Yang; Power, Russell; Zhou, Siyuan; Sovran, Yair; Aguilera, Marcos K.; Li, Jinyang
Abstract:
Currently, users of geo-distributed storage systems face a hard choice between having serializable transactions with high latency, or limited or no transactions with low latency. We show that it is possible to obtain both serializable transactions and low latency, under two conditions. First, transactions are known ahead of time, permitting an a priori static analysis of conflicts. Second, transactions are structured as transaction chains consisting of a sequence of hops, each hop modifying data at one server. To demonstrate this idea, we built Lynx, a geo-distributed storage system that offers transaction chains, secondary indexes, materialized join views, and geo-replication. Lynx uses static analysis to determine if each hop can execute separately while preserving serializability.if so, a client needs wait only for the first hop to complete, which occurs quickly. To evaluate Lynx, we built three applications: an auction service, a Twitter-like microblogging site and a social networking site. These applications successfully use chains to achieve low latency operation and good throughput.
Title: A Note on the Complexity of Model-Checking Bounded Multi-Pushdown Systems
Author(s): Bansal, Kshitij; Demri, Stephane
Abstract:
In this note, we provide complexity characterizations of model checking multi-pushdown systems. Multi-pushdown systems model recursive concurrent programs in which any sequential process has a finite control. We consider three standard notions for boundedness: context boundedness, phase boundedness and stack ordering. The logical formalism is a linear-time temporal logic extending well-known logic CaRet but dedicated to multi-pushdown systems in which abstract operators (related to calls and returns) such as those for next-time and until are parameterized by stacks. We show that the problem is EXPTIME-complete for context-bounded runs and unary encoding of the number of context switches; we also prove that the problem is 2EXPTIME-complete for phase-bounded runs and unary encoding of the number of phase switches. In both cases, the value k is given as an input (whence it is not a constant of the model-checking problem), which makes a substantial difference in the complexity. In certain cases, our results improve previous complexity results.
Title: Learning Hierarchical Feature Extractors For ImageRecognition
Candidate: Boureau, Y-Lan
Advisor(s): LeCun, Yann
Abstract:
Telling cow from sheep is effortless for most animals, but requires much engineering for computers. In this thesis, we seek to tease out basic principles that underlie many recent advances in image recognition. First, we recast many methods into a common unsupervised feature extraction framework based on an alternation of coding steps, which encode the input by comparing it with a collection of reference patterns, and pooling steps, which compute an aggregation statistic summarizing the codes within some region of interest of the image.
Within that framework, we conduct extensive comparative evaluations of many coding or pooling operators proposed in the literature. Our results demonstrate a robust superiority of sparse coding (which decomposes an input as a linear combination of a few visual words) and max pooling (which summarizes a set of inputs by their maximum value). We also propose macrofeatures, which import into the popular spatial pyramid framework the joint encoding of nearby features commonly practiced in neural networks, and obtain significantly improved image recognition performance. Next, we analyze the statistical properties of max pooling that underlie its better performance, through a simple theoretical model of feature activation. We then present results of experiments that confirm many predictions of the model. Beyond the pooling operator itself, an important parameter is the set of pools over which the summary statistic is computed. We propose locality in feature configuration space as a natural criterion for devising better pools. Finally, we propose ways to make coding faster and more powerful through fast convolutional feedforward architectures, and examine how to incorporate supervision into feature extraction schemes. Overall, our experiments offer insights into what makes current systems work so well, and state-of-the-art results on several image recognition benchmarks.
Title: On populations, haplotypes and genome sequencing
Candidate: Franquin, Pierre
Advisor(s): Mishra, Bud
Abstract:
Population genetics has seen a renewed interest since the completion of the human genome project. With the availability of rapidly growing volumes of genomic data, the scientific and medical communities have been optimistic that better understanding of human diseases as well as their treatment were imminent. Many population genomic models and association studies have been designed (or redesigned) to address these problems. For instance, the genome-wide association studies (GWAS) had raised hopes for finding disease markers, personalized medicine and rational drug design. Yet, as of today, they have not yielded results that live up to their promise and have only led to a frustrating disappointment.
Intrigued, but not deterred by these challenges, this dissertation visits the different aspects of these problems. In the first part, we will review the different models and theories of population genetics that are now challenged. We will propose our own implementation of a model to test different hypotheses. This effort will hopefully help us in understanding whether our expectations were unreasonably too high or if we had ignored a crucial piece of information. When discussing association studies, we must not forget that we rely on data that are produced by sequencing technologies, so far available. We have to ensure that the quality of this data is reasonably good for GWAS. Unfortunately, as we will see in the second part, despite the existence of a diverse set of sequencing technologies, none of them can produce haplotypes with phasing, which appears to be the most important type of sequence data needed for association studies. To address this challenge, we propose a novel approach for a sequencing technology, called SMASH that allows us to create the quality and type of haplotypic genome sequences necessary for efficient population genetics.
Title: Optimizing Machine Translation by Learning to Search
Candidate: Galron, Daniel
Advisor(s): Melamed, Dan
Abstract:
We present a novel approach to training discriminative tree-structured machine translation systems by learning to search. We describe three primary innovations in this work: a new parsing coordinator architecture and algorithms to synthesize the required training examples for the learning algorithm; a new semiring that provides an unbiased way to compare translations; and a new training objective that measures whether a translation inference improves the quality of a translation. We also apply the reinforcement learning concept of exploration to SMT. Finally, we empirically evaluate the effects of our innovations on the quality of translations output by our system.
Title: Flexible-Cost SLAM
Candidate: Grimes, Matthew
Advisor(s): LeCun, Yann
Abstract:
The ability of a robot to track its position and its surroundings is critical in mobile robotics applications, such as autonomous transport, farming, search-and-rescue, and planetary exploration.
As a foundational building block to such tasks, localization must remain reliable and unobtrusive. For example, it must not provide an unneeded level of precision, when the cost of doing so displaces higher-level tasks from a busy CPU. Nor should it produce noisy estimates on the cheap, when there are CPU cycles to spare.
This thesis explores localization solutions that provide exactly the amount of accuracy needed to a given task. We begin with a real-world system used in the DARPA Learning Applied to Ground Robotics (LAGR) competition. Using a novel hybrid of wheel and visual odometry, we cut the cost of visual odometry from 100% of a CPU to 5%, clearing room for other critical visual processes, such as long-range terrain classification. We present our hybrid odometer in chapter 2.
Next, we describe a novel SLAM algorithm that provides a means to choose the desired balance between cost and accuracy. At its fastest setting, our algorithm converges faster than previous stochastic SLAM solvers, while maintaining significantly better accuracy. At its most accurate, it provides the same solution as exact SLAM solvers. Its main feature, however, is the ability to flexibly choose any point between these two extremes of speed and precision, as circumstances demand. As a result, we are able to guarantee real-time performance at each timestep on city-scale maps with large loops. We present this solver in chapter 3, along with results from both commonly available datasets and Google Street View data.
Taken as a whole, this thesis recognizes that precision and efficiency can be competing values, whose proper balance depends on the application and its fluctuating circumstances. It demonstrates how a localizer can and should fit its cost to the task at hand, rather than the other way around. In enabling this flexibility, we demonstrate a new direction for SLAM research, as well as provide a new convenience for end-users, who may wish to map the world without stopping it.
Title: SMT Beyond DPLL(T): A New Approach to Theory Solvers and Theory Combination
Candidate: Jovanovic, Dejan
Advisor(s): Barrett, Clark
Abstract:
Satisifiability modulo theories (SMT) is the problem of deciding whether a given logical formula can be satisifed with respect to a combination of background theories. The past few decades have seen many significant developments in the field, including fast Boolean satisfiability solvers (SAT), efficient decision procedures for a growing number of expressive theories, and frameworks for modular combination of decision procedures. All these improvements, with addition of robust SMT solver implementations, culminated with the acceptance of SMT as a standard tool in the fields of automated reasoning and computer aided verification. In this thesis we develop new decision procedures for the theory of linear integer arithmetic and the theory of non-linear real arithmetic, and develop a new general framework fro combination of decision procedures. The new decision procedures integrate theory specific reasoning and the Boolean search to provide more powerful and efficient procedures, and allow a more expressive language for explaining problematic states. The new framework for combination of decision procedures overcomes the complexity limitations and restrictions on the theories imposed by the standard Nelson-Oppen approach.
Title: An Adaptive Fast Multipole Method-Based PDE Solver in Three Dimensions
Candidate: Langston, Matthew Harper
Advisor(s): Zorin, Denis
Abstract:
Many problems in scientific computing require the accurate and fast solution to a variety of elliptic PDEs. These problems become increasingly dif.cult in three dimensions when forces become non-homogeneously distributed and geometries are complex.
We present an adaptive fast volume solver using a new version of the fast multipole method, incorporated with a pre-existing boundary integral formulation for the development of an adaptive embedded boundary solver.
For the fast volume solver portion of the algorithm, we present a kernel-independent, adaptive fast multipole method of arbitrary order accuracy for solving elliptic PDEs in three dimensions with radiation boundary conditions. The algorithm requires only a Greenâs function evaluation routine for the governing equation and a representation of the source distribution (the right-hand side) that can be evaluated at arbiÂtrary points.
The performance of the method is accelerated in two ways. First, we construct a piecewise polynomial approximation of the right-hand side and compute far-.eld expansions in the FMM from the coef.cients of this approximation. Second, we precompute tables of quadratures to handle the near-.eld interactions on adaptive octree data structures, keeping the total storage requirements in check through the exploitation of symmetries. We additionally show how we extend the free-space volume solver to solvers with periodic and well as Dirichlet boundary conditions.
For incorporation with the boundary integral solver, we develop interpolation methods to maintain the accuracy of the volume solver. These methods use the existing FMM-based octree structure to locate apÂpropriate interpolation points, building polynomial approximations to this larger set of forces and evaluating these polynomials to the locally under-re.ned grid in the area of interest.
We present numerical examples for the Laplace, modi.ed Helmholtz and Stokes equations for a variety of boundary conditions and geometries as well as studies of the interpolation procedures and stability of far-.eld and polynomial constructions.
Title: Acquiring information from wider scope to improve event extraction
Candidate: Liao, Shasha
Advisor(s): Grishman, Ralph
Abstract:
Event extraction is a particularly challenging type of information extraction (IE). Most current event extraction systems rely on local information at the phrase or sentence level. However, this local context may be insufficient to resolve ambiguities in identifying particular types of events; information from a wider scope can serve to resolve some of these ambiguities.
In this thesis, we first investigate how to extract supervised and unsupervised features to improve a supervised baseline system. Then, we present two additional tasks to show the benefit of wider scope features in semi-supervised learning (self-training) and active learning (co-testing). Experiments show that using features from wider scope can not only aid a supervised local event extraction baseline system, but also help the semi-supervised or active learning approach.
Title: A tool for extracting and indexing spatio-temporal information from biographical articles in Wikipedia
Candidate: Morton-Owens, Emily
Advisor(s): Davis, Ernest
Abstract:
The Kivrin program, consisting of a crawler, a data collection, and a front-end interface, attempts to extract biographical information from Wikipedia, specifically, spatio-temporal information--who was where when--and make it easily searchable. Some of the considerations standard to moving object databases do not apply in this context, because the texts by their nature discuss a discontinuous series of notable moments. The paper discusses different methods of arranging the crawler queue priority to find more important figures and of disambiguating locations when the same place name (toponym) is shared among several places. When lifespan information is not available, it is estimated to exclude sightings outside the person's plausible lifetime.
The results are grouped by the number of sightings in the user's search range to minimize the visibility of false drops when they occur. Erroneous results are more visible in times and places where fewer legitimate sightings are recorded; the data is skewed, like Wikipedia itself, towards the U.S. and Western Europe and relatively recent history. The system could be most improved by using statistical methods to predict which terms are more likely personal names than place names and to identify verbs that precede location information rather than personal names. It could also be improved by incorporating the times as a third dimension in the geospatial index, which would allow "near" queries to include that dimension rather than a strict range.
The program can be used at http://linserv1.cims.nyu.edu:48866/cgi-bin/index.cgi
Title: Mobile Accessibility Tools for the Visually Impaired
Candidate: Paisios, Nektarios
Advisor(s): Subramanian; Lakshminarayanan
Abstract:
Visually impaired users are in dire need of better accessibility tools. The past few years have witnessed an exponential growth in the computing capabilities and onboard sensing capabilities of mobile phones making them an ideal candidate for building next-generation applications. We believe that the mobile device can play a significant role in the future for aiding visually impaired users in day-to-day activities with simple and usable mobile accessibility tools. This thesis describes the design, implementation, evaluation and user-study based analysis of four different mobile accessibility applications.
Our first system is the design of a highly accurate and usable mobile navigational guide that uses Wi-Fi and accelerometer sensors to navigate unfamiliar environments. A visually impaired user can use the system to construct a virtual topological map across points of interest within a building based on correlating the user' walking patterns (with turn signals) with the Wi-Fi and accelerometer readings. The user can subsequently use the map to navigate previously traveled routes. Our second system, Mobile Brailler, presents several prototype methods of text entry on a modern touch screen mobile phone that are based on the Braille alphabet and thus are convenient for visually impaired users. Our third system enables visually impaired users to leverage the camera of a mobile device to accurately recognize currency bills even if the images are partially or highly distorted. The final system enables visually impaired users to determine whether a pair of clothes, in this case of a tie and a shirt, can be worn together or not, based on the current social norms of color-matching.
We believe that these applications together, provide a suite of important mobile accessibility tools to enhance four critical aspects of a day-to-day routine of a visually impaired user: to navigate easily, to type easily, to recognize currency bills (for payments) and to identify matching clothes.
Title: Reusable Software Infrastructure for Stream Processing
Candidate: Soule, Robert
Advisor(s): Grimm, Robert
Abstract:
Developers increasingly use streaming languages to write their data processing applications. While a variety of streaming languages exist, each targeting a particular application domain, they are all similar in that they represent a program as a graph of streams (i.e. sequences of data items) and operators (i.e. data transformers). They are also similar in that they must process large volumes of data with high throughput. To meet this requirement, compilers of streaming languages must provide a variety of streaming-specific optimizations, including automatic parallelization. Traditionally, when many languages share a set of optimizations, language implementors translate the source languages into a common representation called an intermediate language (IL). Because optimizations can modify the IL directly, they can be re-used by all of the source languages, reducing the overall engineering effort. However, traditional ILs and their associated optimizations target single-machine, single-process programs. In contrast, the kinds of optimizations that compilers must perform in the streaming domain are quite different, and often involve reasoning across multiple machines. Consequently, existing ILs are not suited to streaming languages.
This thesis addresses the problem of how to provide a reusable infrastructure for stream processing languages. Central to the approach is the design of an intermediate language specifically for streaming languages and optimizations. The hypothesis is that an intermediate language designed to meet the requirements of stream processing can assure implementation correctness; reduce overall implementation effort; and serve as a common substrate for critical optimizations. In evidence, this thesis provides the following contributions: (1) a catalog of common streaming optimizations that helps define the requirements of a streaming IL; (2) a calculus that enables reasoning about the correctness of source language translation and streaming optimizations; and (3) an intermediate language that preserves the semantics of the calculus, while addressing the implementation issues omitted from the calculus This work significantly reduces the effort it takes to develop stream processing languages, and jump-starts innovation in language and optimization design.
Title: Hitting the Sweet Spot for Streaming Languages: Dynamic Expressivity with Static Optimization
Author(s): Soulé, Robert; Gordon, Michael I.; Amarasinghe, Saman; Grimm, Robert; Hirzel, Martin
Abstract:
Developers increasingly use stream processing languages to write applications that process large volumes of data with high throughput. Unfortunately, when choosing which stream processing language to use, they face a difficult choice. On the one hand, dynamically scheduled languages allow developers to write a wider range of applications, but cannot take advantage of many crucial optimizations. On the other hand, statically scheduled languages are extremely performant, but cannot express many important streaming applications.
This paper presents the design of a hybrid scheduler for stream processing languages. The compiler partitions the streaming application into coarse-grained subgraphs separated by dynamic rate boundaries. It then applies static optimizations to those subgraphs. We have implemented this scheduler as an extension to the StreamIt compiler, and evaluated its performance against three scheduling techniques used by dynamic systems: OS thread, demand, and no-op. Our scheduler not only allows the previously static version of StreamIt to run dynamic rate applications, but it outperforms the three dynamic alternatives. This demonstrates that our scheduler strikes the right balance between expressivity and performance for stream processing languages.
Title: Building scalable geo-replicated storage backends for web applications
Candidate: Sovran, Yair
Advisor(s): Li, Jinyang
Abstract:
Web applications increasingly require a storage system that is both scalable and can replicate data across many distant data centers or sites. Most existing storage solutions fall into one of two categories: Traditional databases offer strict consistency guarantees and programming ease, but are difficult to scale in a geo-replicated setting. NoSQL stores are scalable and efficient, but have weak consistency guarantees, placing the burden of ensuring consistency on programmers. In this dissertation, we describe two systems that help bridge the two extremes, providing scalable, geo-replicated storage for web applications, while also easy to program for. Walter is a key-value store that supports transactions and replicating data across distant sites. A key feature underlying Walter is a new isolation property: Parallel Snapshot Isolation (PSI). PSI allows Walter to replicate data asynchronously, while providing strong guarantees within each site. PSI does not allow write-write conflicts, alleviating the burden of writing conflict resolution logic. To prevent write-write conflicts and implement PSI, Walter uses two new and simple techniques: preferred sites and counting sets. Lynx is a distributed database backend for scaling latency-sensitive web applications. Lynx supports optimizing queries via data denormalization, distributed secondary indexes, and materialized join views. To preserve data constraints across denormalized tables and secondary indexes, Lynx relies on the a novel primitive: Distributed Transaction Chain (DTC). A DTC groups a sequence of transactions to be executed on different nodes while providing two guarantees. First, all transactions in a DTC execute exactly once despite failures. Second, transactions from concurrent DTCs are interleaved consistently on common nodes. We built several web applications on top of Walter and Lynx: an auction service, a microblogging service, and a social networking website. We have found that building web applications using Walter and Lynx is quick and easy. Our experiments show that the resulting applications are capable of providing scalable, low latency operation across multiple geo-replicated sites.
Title: Rapid Training of Information Extraction with Local and Global Data Views
Candidate: Sun, Ang
Advisor(s): Grishman, Ralph
Abstract:
This dissertation focuses on fast system development for Information Extraction (IE). State-of-the-art systems heavily rely on extensively annotated corpora, which are slow to build for a new domain or task. Moreover, previous systems are mostly built with local evidence such as words in a short context window or features that are extracted at the sentence level. They usually generalize poorly on new domains.
This dissertation presents novel approaches for rapidly training an IE system for a new domain or task based on both local and global evidence. Specifically, we present three systems: a relation type extension system based on active learning, a relation type extension system based on semi-supervised learning, and a cross-domain bootstrapping system for domain adaptive named entity extraction.
The active learning procedure adopts features extracted at the sentence level as the local view and distributional similarities between relational phrases as the global view. It builds two classifiers based on these two views to find the most informative contention data points to request human labels so as to reduce annotation cost.
The semi-supervised system aims to learn a large set of accurate patterns for extracting relations between names from only a few seed patterns. It estimates the confidence of a name pair both locally and globally: locally by looking at the patterns that connect the pair in isolation; globally by incorporating the evidence from the clusters of patterns that connect the pair. The use of pattern clusters can prevent semantic drift and contribute to a natural stopping criterion for semi-supervised relation pattern discovery.
For adapting a named entity recognition system to a new domain, we propose a cross-domain bootstrapping algorithm, which iteratively learns a model for the new domain with labeled data from the original domain and unlabeled data from the new domain. We first use word clusters as global evidence to generalize features that are extracted from a local context window. We then select self-learned instances as additional training examples using multiple criteria, including some based on global evidence.
Title: Combating Sybil attacks in cooperative systems
Candidate: Tran, Nguyen
Advisor(s): Li, Jinyang
Abstract:
Cooperative systems are ubiquitous nowadays. In a cooperative system, end users contribute resource to run the service instead of only receiving the service passively from the system. For example, users upload and comment pictures and videos on Flicker and YouTube, users submit and vote on news articles on Digg. As another example, users in BitTorrent contribute bandwidth and storage to help each other download content. As long as users behave as expected, these systems benefit immensely from user contribution. In fact, five out of ten most popular websites are operating in this cooperative fashion (Facebook, YouTube, Blogger, Twitter, Wikipedia). BitTorrent is dominating the global Internet traffic.
A robust cooperative system cannot blindly trust that its users will truthfully participate in the system. Malicious users seek to exploit the systems for profit. Selfish users consume but avoid to contribute resource. For example, adversaries have manipulated the voting system of Digg to promote their articles of dubious quality. Selfish users in public BitTorrent communities leave the system to avoid uploading files to others, resulting in drastic performance degradation for these content distribution systems. The ultimate way to disrupt security and incentive mechanisms of cooperative systems is using Sybil attacks, in which the adversary creates many Sybil identities (fake identities) and use them to disrupt the systems' normal operation. No security and incentive mechanism works correctly if the systems do not have a robust identity management that can defend against Sybil attacks.
This thesis provides robust identity management schemes which are resilient to the Sybil attack, and use them to secure and incentivize user contribution in several example cooperative systems. The main theme of this work is to leverage the social network among users in designing secure and incentive-compatible cooperative systems. First, we develop a distributed admission control protocol, called Gatekeeper, that leverages social network to admit most honest user identities and only few Sybil identities into the systems. Gatekeeper can be used as a robust identity management for both centralized and decentralized cooperative systems. Second, we provide a vote aggregation system for content voting systems, called SumUp, that can prevent an adversary from casting many bogus votes for a piece of content using the Sybil attack. SumUp leverages unique properties of content voting systems to provide significantly better Sybil defense compared with applying a general admission control protocol such as \gatekeeper. Finally, we provide a robust reputation system, called Credo, that can be used to incentivize bandwidth contribution in peer-to-peer content distribution networks. Credo reputation can capture user contribution, and is resilient to both Sybil and collusion attacks.
Title: Multi-species biclustering: An integrative method to identify functional gene conservation between multiple species
Candidate: Waltman, Peter
Advisor(s): Bonneau, Richard
Abstract:
Background : Several recent comparative functional genomics projects have indicated that the co-regulation of many genes is conserved across species, at least in part. This suggests that comparative analysis of functional genomics data-sets could prove powerful in identifying co-regulated groups that are conserved across multiple species.
Results : We present recent work to extend our cMonkey algorithm to simultaneously bicluster heterogeneous data from multiple species to identify conserved modules of orthologous genes, which can yield evolutionary insights into the formation of regulatory modules. We also present results from the multi-species analysis to two triplets of bacteria. The first of these is a triplet of Gram-positive bacteria consisting of Bacillus subtilis, Bacillus anthracis, and Listeria monocytogenes, while the second is a triplet of Gram-negative bacteria that includes Escherichia coli, Salmonella typhimurium and Vibrio cholerae. Finally, we will present initial results from the multi-species biclustering analysis of human and mouse hematopoietic differentiation data.
Conclusion : Analysis of biclusters obtained revealed a surprising number of gene groups with conserved modularity and high biological significance as judged by several measures of cluster quality. We also highlight cases of interest from the Gram-positive triplet, including one that suggests a temporal difference in the expression of genes governing sporulation in the two Bacillus species. While analysis of the mouse and human hematopoietic differentiation is preliminary, it indicates the applicability of this analysis to eukaryotic systems, including comparison of cancer model systems. Finally, we suggest ways in which this analysis could be extended to identify divergent modules that may exist between normal and disease tissue.
Title: Collusion Preserving Computation
Candidate: Alwen, Joel
Advisor(s): Dodis, Yevgeniy
Abstract:
In collusion-free protocols, subliminal communication is impossible and parties are thus unable to communicate any information beyond what the protocol allows". Collusion-free protocols are interesting for several reasons, but have specifically attracted attention because they can be used to reduce trust in game-theoretic mechanisms. Collusion-free protocols are impossible to achieve (in general) when all parties are connected by point-to-point channels, but exist under certain physical assumptions (Lepinksi et al., STOC 2005) or in specific network topologies (Alwen et al., Crypto 2008).
In addition to proposing the definition, we explore necessary properties of the underlying communication resource. Next we provide a general feasibility result for collusion-preserving computation of arbitrary functionalities. We show that the resulting protocols enjoy an elegant (and surprisingly strong) fallback security even in the case when the underlying communication resource acts in a Byzantine manner. Finally, we investigate the implications of these results in the context of mechanism design.
Title: Re-architecting Web and Mobile Information Access for Emerging Regions
Candidate: Chen, Jay
Advisor(s): Subramanian; Lakshminarayanan
Abstract:
Providing access to information for people in emerging regions is an important problem. Over the past decade there have been many proposed and increasingly numerous deployed systems to enable information access, but successes are few and modest at best. Internet in emerging regions is still generally unusable or intolerably slow. Mobile phone applications are either not designed for the phones that poor people own, otherwise, the applications lack functionality, are difficult to use, or expensive to operate. In this work we focus on enabling digital information access for people in emerging regions.
To advance the state of the art, we contribute numerous observations about how people access information in emerging regions, why the current models for web access and SMS platforms are broken, and techniques to enable applications over constrained Internet or SMS. The mechanisms presented here were designed after extensive field work in several different regions including rural, peri-urban, and urban areas in India, Kenya, Ghana, and Mexico. Multiple user studies were conducted throughout the course of system design and prototyping. We present a novel set of context appropriate platforms and tools, some spanning several layers of the networking stack. Five complete systems were implemented and deployed in the field. First, Event Logger for Firefox (ELF) is an easily deployable Firefox extension which functions as both a web browsing analysis tool and an in-browser web optimization platform. Second, RuralCafe provides a platform for web search and browsing over extremely slow or intermittent networks. Third, Contextual Information Portals (CIP) provide cached repositories of web pages tailored to the particular context in which it is to be used. Fourth, UjU is a mobile application platform that simplies the design of new SMS-based mobile applications. Finally, SMSFind is a SMS-based search service that runs on mobile phones without setup or subscription to a data plan.
Taken as a whole, the systems here are a comprehensive solution for addressing the problem of enabling digital information access in emerging regions.
Title: Automatic Deduction for Theories of Algebraic Data Types
Candidate: Chikanian, Igor
Advisor(s): Barrett, Clark
Abstract:
In this thesis we present formal logical systems, concerned with reasoning about algebraic data types.
The first formal system is based on the quantifier-free calculus (outermost universally quantified). This calculus is comprised of state change rules, and computations are performed by successive applications of these rules. Thereby, our calculus gives rise to an abstract decision procedure. This decision procedure determines if a given formula involving algebraic type members is valid. It is shown that this calculus is sound and complete. We also examine how this system performs practically and give experimental results. Our main contribution, as compared to previous work on this subject,is a new and more efficient decision procedure for checking satisfiability of the universal fragment within the theory of algebraic data types.
The second formal system, called Term Builder, is the deductive system based on higher order type theory, which subsumes second order and higher order logics. The main purpose of this calculus is to formulate and prove theorems about algebraic or other arbitrary user-defined types.Term Builder supports proof objects and is both, an interactive theorem prover, and verifier. We describe the built-in deductive capabilities of Term Builder and show its consistency. The logic represented by our prover is intuitionistic. Naturally, it is also incomplete and undecidable, but its expressive power is much higher than that of the first formal system.
Among our achievements in building this theorem prover is an elegant and intuitive GUI for building proofs. Also, a new feature from the foundational viewpoint is that, in contrast with other approaches, we have uniqueness-of-types property, which is not modulo beta-conversion.
Title: Two-Level Overlapping Schwarz Algorithms for a Staggered Discontinuous Galerkin Method
Author(s): Chung, Eric T.; Kim, Hyea Hyun; Widlund, Olof B.
Abstract:
Two overlapping Schwarz algorithms are developed for a discontinuous Galerkin (DG) finite element approximation of second order scalar elliptic problems in both two and three dimensions. The discontinuous Galerkin formulation is based on a staggered discretization introduced by Chung and Engquist for the acoustic wave equation. Two types of coarse problems are introduced for the two-level Schwarz algorithms. The first is built on a non-overlapping subdomain partition, which allows quite general subdomain partitions, and the second on introducing an additional coarse triangulation that can also be quite independent of the fine triangulation. Condition number ounds are established and numerical results are presented.
Title: An Alternative Coarse Space for Irregular Subdomains and an Overlapping Schwarz Algorithm
Author(s): Dohrmann, Clark R.; Widlund, Olof B.
Abstract:
In earlier work on domain decomposition methods for elliptic problems in the plane, an assumption that each subdomain is triangular, or a union of a few coarse triangles, has often been made. This is similar to what is required in geometric multigrid theory and is unrealistic if the subdomains are produced by a mesh partitioner. In an earlier paper, coauthored with Axel Klawonn, the authors introduced a coarse subspace for an overlapping Schwarz method with one degree of freedom for each subdomain vertex and one for each subdomain edge. A condition number bound proportional to $(1+\log(H/h))^2(1+H/\delta)$ was established assuming only that the subdomains are John domains; here $H/\delta$ measures the relative overlap between neighboring subdomains and $H/h$ the maximum number of elements across individual subdomains. We were also able to relate the rate of convergence to a parameter in an isoperimetric inequality for the subdomains into which the domain of the problem has been partitioned.
In this paper, the dimension of the coarse subspace is decreased by using only one degree of freedom for each subdomain vertex; if all subdomains have three edges, this leads to a reduction of the dimension of the coarse subspace by approximately a factor four. In addition, the condition number bound is shown to be proportional to $(1+\log(H/h))(1+H/\delta)$ under a quite mild assumption on the relative length of adjacent subdomain edges.
In this study, the subdomains are assumed to be uniform in the sense of Peter Jones. As in our earlier work, the results are insensitive to arbitrary large jumps in the coefficients of the elliptic problem across the interface between the subdomains.
Numerical results are presented which confirm the theory and demonstrate the usefulness of the algorithm for a variety of mesh decompositions and distributions of material properties. It is also shown that the new algorithm often converges faster than the older one in spite of the fact that the dimension of the coarse space has been decreased considerably.
Title: Parsing All of C by Taming the Preprocessor
Author(s): Gazzillo, Paul; Grimm, Robert
Abstract:
Given the continuing popularity of C for building large-scale programs, such as Linux, Apache, and Bind, it is critical to provide effective tool support, including, for example, code browsing, bug finding, and automated refactoring. Common to all such tools is a need to parse C. But C programs contain not only the C language proper but also preprocessor invocations for file inclusion (#include), conditional compilation (#if, #ifdef, and so on), and macro definition/expansion (#define). Worse, the preprocessor is a textual substitution system, which is oblivious to C constructs and operates on individual tokens. At the same time, the preprocessor is indispensable for improving C's expressivity, abstracting over software/hardware dependencies, and deriving variations from the same code base. The x86 version of the Linux kernel, for example, depends on about 7,600 header files for file inclusion, 7,000 configuration variables for conditional compilation, and 520,000 macros for code expansion.
In this paper, we present a new tool for parsing all of C, including arbitrary preprocessor use. Our tool, which is called SuperC, is based on a systematic analysis of all interactions between lexing, preprocessing, and parsing to ensure completeness. It first lexes and preprocesses source code while preserving conditionals. It then parses the result using a novel variant of LR parsing, which automatically forks parsers when encountering a conditional and merges them again when reaching the same input in the same state. The result is a well-formed AST, containing static choice nodes for conditionals. While the parsing algorithm and engine are new, neither grammar nor LR parser table generator need to change. We discuss the results of our problem analysis, the parsing algorithm itself, the pragmatics of building a real-world tool, and a demonstration on the x86 version of the Linux kernel.
Title: Efficient Cryptographic Primitives for Non-Interactive Zero-Knowledge Proofs and Applications
Candidate: Haralambiev, Kristiyan
Advisor(s): Shoup, Victor
Abstract:
Non-interactive zero-knowledge (NIZK) proofs have enjoyed much interest in cryptography since they were introduced more than twenty years ago by Blum et al. [BFM88]. While quite useful when designing modular cryptographic schemes, until recently NIZK could be realized efficiently only using certain heuristics. However, such heuristic schemes have been widely criticized. In this work we focus on designing schemes which avoid them. In [GS08], Groth and Sahai presented the first efficient (and currently the only) NIZK proof system in the standard model. The construction is based on bilinear maps and is limited to languages of certain satisfiable system of equations. Given this expressibility limitation of the system of equations, we are interested in cryptographic primitives that are "compatible" with it. Equipped with such primitives and Groth-Sahai proof system, we show how to construct cryptographic schemes efficiently in a modular fashion.
In this work, we describe properties required by any cryptographic scheme to mesh well with Groth-Sahai proofs. Towards this, we introduce the notion of "structure-preserving" cryptographic scheme. We present the first constant-size structure-preserving signature scheme for messages consisting of general bilinear group elements. This allows us (for the first time) to instantiate efficiently a modular construction of round-optimal blind signature based on the framework of Fischlin [Fis06].
Our structure-preserving homomorphic trapdoor commitment schemes yield efficient leakage-resilient signatures (in the bounded leakage model) which satisfy the standard security requirements and additionally tolerates any amount of leakage; all previous works satisfied at most two of those three properties.
Lastly, we build a structure-preserving encryption scheme which satisfies the standard CCA security requirements. While somewhat similar to the notion of verifiable encryption, it provides better properties and yields the first efficient two-party protocol for joint ciphertext computation. Note that the efficient realization of such a protocol was not previously possible even using the heuristics mentioned above.
Lastly, in this line of work, we revisit the notion of simulation extractability and define "true-simulation extractable" NIZK proofs. Although quite similar to the notion of simulation-sound extractable NIZK proofs, there is a subtle but rather important difference which makes it weaker and easier to instantiate efficiently. As it turns out, in many scenarios, this new notion is sufficient, and using it, we can construct efficient leakage resilient signatures and CCA encryption scheme.
Title: Sharing is Caring: Combination of Theories
Author(s): Jovanovic, Dejan; Barrett, Clark
Abstract:
One of the main shortcomings of the traditional methods for combining theories is the complexity of guessing the arrangement of the variables shared by the individual theories. This paper presents a reformulation of the Nelson-Oppen method that takes into account explicit equality propagation and can ignore pairs of shared variables that the theories do not care about. We show the correctness of the new approach and present care functions for the theory of uninterpreted functions and the theory of arrays. The effectiveness of the new method is illustrated by experimental results demonstrating a dramatic performance improvement on benchmarks combining arrays and bit-vectors.
Title: Learning Feature Hierarchies for Object Recognition
Candidate: Kavukcuoglu, Koray
Advisor(s): LeCun, Yann
Abstract:
In this thesis we study unsupervised learning algorithms for training feature extractors and building deep learning models. We propose sparse-modeling algo- rithms as the foundation for unsupervised feature extraction systems. To reduce the cost of the inference process required to obtain the optimal sparse code, we model a feed-forward function that is trained to predict this optimal sparse code. Using an efficient predictor function enables the use of sparse coding in hierarchical models for object recognition. We demonstrate the performance of the developed system on several recognition tasks, including object recognition, handwritten digit classification and pedestrian detection. Robustness to noise or small variations in the input is a very desirable property for a feature extraction algorithm. In order to train locally-invariant feature extractors in an unsupervised manner, we use group sparsity criteria that promote similarity between the dictionary elements within a group. This model produces locally-invariant representations under small pertur- bations of the input, thus improving the robustness of the features. Many sparse modeling algorithms are trained on small image patches that are the same size as the dictionary elements. This forces the system to learn multiple shifted versions of each dictionary element. However, when used convolutionally over large im- ages to extract features, these models produce very redundant representations. To avoid this problem, we propose convolutional sparse coding algorithms that yield a richer set of dictionary elements, reduce the redundancy of the representation and improve recognition performance.
Title: Topics in Formal Synthesis and Modeling
Candidate: Klein, Uri
Advisor(s): Pnueli, Amir; Zuck, Lenore
Abstract:
The work presented focuses on two problems, that of synthesizing systems from formal specifications, and that of formalizing REST -- a popular web applications' development pattern.
For the synthesis problem, we distinguish between the synchronous and the asynchronous case. For the former, we solve a problem concerning a fundamental flaw in specification construction in previous work. We continue with exploring effective synthesis of asynchronous systems (programs on multi-threaded systems). Two alternative models of asynchrony are presented, and shown to be equally expressive for the purpose of synthesis.
REST is a software architectural style used for the design of highly scalable web applications. Interest in REST has grown rapidly over the past decade. However, there is also considerable confusion surrounding REST: many examples of supposedly RESTful APIs violate key REST constraints. We show that the constraints of REST and of RESTful HTTP can be precisely formulated within temporal logic. This leads to methods for model checking and run-time verification of RESTful behavior. We formulate several relevant verification questions and analyze their complexity.
Title: Effective Synthesis of Asynchronous Systems from GR(1) Specifications
Author(s): Klein, Uri; Piterman, Nir; Pnueli, Amir
Abstract:
We consider automatic synthesis from linear temporal logic specifications for asynchronous systems. We aim the produced reactive systems to be used as software in a multi-threaded environment. We extend previous reduction of asynchronous synthesis to synchronous synthesis to the setting of multiple input and multiple output variables. Much like synthesis for synchronous designs, this solution is not practical as it requires determinization of automata on infinite words and solution of complicated games. We follow advances in synthesis of synchronous designs, which restrict the handled specifications but achieve scalability and efficiency. We propose a heuristic that, in some cases, maintains scalability for asynchronous synthesis. Our heuristic can prove that specifications are realizable and extract designs. This is done by a reduction to synchronous synthesis that is inspired by the theoretical reduction.
Title: Formalization and Automated Verification of RESTful Behavior
Author(s): Klein, Uri; Namjoshi, Kedar S.
Abstract:
REST is a software architectural style used for the design of highly scalable web applications. Interest in REST has grown rapidly over the past decade, spurred by the growth of open web APIs. On the other hand, there is also considerable confusion surrounding REST: many examples of supposedly RESTful APIs violate key REST constraints. We show that the constraints of REST and of RESTful HTTP can be pre- cisely formulated within temporal logic. This leads to methods for model checking and run-time verfication of RESTful behavior. We formulate several relevant verification questions and analyze their complexity.
Title: Domain Decomposition Methods for Reissner-Mindlin Plates Discretized with the Falk-Tu Elements
Author(s): Lee, Jong Ho
Abstract:
The Reissner-Mindlin plate theory models a thin plate with thickness t. The condition number of finite element approximations of this model deteriorates badly as the thickness t of the plate converges to 0. In this thesis, we develop an overlapping domain decomposition method for the Reissner-Mindlin plate model discretized by Falk-Tu elements with a convergence rate which does not deteriorate when t converges to 0. We use modern overlapping methods which use the Schur complements to define coarse basis functions and show that the condition number of this overlapping method is bounded by C(1 + H/delta )^3*(1 + log(H/h))^2. Here H is the maximum diameter of the subdomains, delta the size of overlap between subdomains, and h the element size. Numerical examples are provided to confirm the theory. We also modify the overlapping method to develop a BDDC method for the Reissner-Mindlin model. We establish numerically an extension lemma to obtain a constant bound and an edge lemma to obtain a C(1 + log(H/h))^2 bound. Given such bounds, the condition number of this BDDC method is shown to be bounded by C(1 + log(H/h))^2.
Title: Adaptive Isotopic Approximation of Nonsingular Curves and Surfaces
Candidate: Lin, Long
Advisor(s): Yap, Chee
Abstract:
Consider the problem of computing isotopic approximations of nonsingular curves and surfaces that are implicitly represented by equations of the form f (X, Y )=0 and f (X,Y, Z)=0. Thisfundamentalproblem has seen much progress along several fronts, but we will focus on domain subdivision algorithms. Two algorithms in this area are from Snyder(1992) and Plantinga and Vegter(2004). We introduce a family of new algorithms that combines the advantages of these two algorithms: like Snyder, we use the parameterizability criterion for subdivision, and like Plantinga and Vegter, we exploit nonlocal isotopy.
We first apply our approach to curves, resulting in a more efficient algorithm. We then extend our approach to surfaces. The extension is by no means routine, as the correctness arguments and case analysis are more subtle. Also, a new phenomenon arises in which local rules for constructing surfaces are no longer sufficient.
We further extend our algorithms in two important and practical directions: first, we allow subdivision cells to be non squares or non cubes, with arbitrary but bounded aspect ratios: in 2D, we allow boxes to be split into 2 or 4 children; and in 3D, we allow boxes to be split into 2, 4 or 8 children. Second, we allow the inputregion-of-interest(ROI) to have arbitrary geometry represented by anquadtreeoroctree,aslongas the curves or surfaces has no singularities in the ROI and intersects the boundary of ROI transversally.
Our algorithm is numerical because our primitives are based on interval arithmetic and exact BigFloat numbers. It is practical, easy to implement exactly (compared to algebraic approaches) and does not suffer from implementation gaps (compared to geometric approaches). We report some very encouraging experimental results,showing that our algorithms can be much more efficient than the algorithms of Plantinga and Vegter(2D and 3D)and Snyder(2D only).
Title: Real-Space Localization Methods for Minimizing the Kohn-Sham Energy
Candidate: Millstone, Marc
Advisor(s): Overton, Michael
Abstract:
The combination of ever increasing computational power and new mathematical models has fundamentally changed the field of computational chemistry. One example of this is the use of new algorithms for computing the charge density of a molecular system from which one can predict many physical properties of the system.
This thesis presents two new algorithms for minimizing the Kohn-Sham energy, which is used to describe a system of non-interacting electrons through a set of single-particle wavefunctions. By exploiting a known localization region of the wavefunctions, each algorithm evaluates the Kohn-Sham energy function and gradient at a set of iterates that have a special sparsity structure. We have chosen to represent the problem in real-space using finite-differences, allowing us to efficiently evaluate the energy function and gradient using sparse linear algebra. Detailed numerical experiments are provided on a set of representative molecules demonstrating the performance and robustness of these methods.
Title: Scoring-and-Unfolding Trimmed Tree Assembler: Algorithms for Assembling Genome Sequences Accurately and Efficiently
Candidate: Narzisi, Giuseppe
Advisor(s): Mishra, Bud
Abstract:
The recent advances in DNA sequencing technology and their many potential applications to Biology and Medicine have rekindled enormous interest in several classical algorithmic problems at the core of Genomics and Computational Biology: primarily, the whole-genome sequence assembly problem (WGSA). Two decades back, in the context of the Human Genome Project, the problem had received unprecedented scientific prominence: its computational complexity and intractability were thought to have been well understood; various competitive heuristics, thoroughly explored and the necessary software, properly implemented and validated. However, several recent studies, focusing on the experimental validation of de novo assemblies, have highlighted several limitations of the current assemblers.
Intrigued by these negative results, this dissertation reinvestigates the algorithmic techniques required to correctly and efficiently assemble genomes. Mired by its connection to a well-known NP-complete combinatorial optimization problem, historically, WGSA has been assumed to be amenable only to greedy and heuristic methods. By placing efficiency as their priority, these methods opted to rely on local searches, and are thus inherently approximate, ambiguous or error-prone. This dissertation presents a novel sequence assembler, SUTTA, that dispenses with the idea of limiting the solutions to just the approximated ones, and instead favors an approach that could potentially lead to an exhaustive (exponential-time) search of all possible layouts but tames the complexity through constrained search (Branch-and-Bound) and quick identification and pruning of implausible solutions.
Complementary to this problem is the task of validating the generated assemblies. Unfortunately, no commonly accepted method exists yet and widely used metrics to compare the assembled sequences emphasize only size, poorly capturing quality and accuracy. This dissertation also addresses these concerns by developing a more comprehensive metric, the Feature-Response Curve, that, using ideas from classical ROC (receiver-operating characteristic) curve, more faithfully captures the trade-off between contiguity and quality.
Finally, this dissertation demonstrates the advantages of a complete pipeline integrating base-calling (TotalReCaller) with assembly (SUTTA) in a Bayesian manner.
Title: Domain Decomposition Methods for Raviart-Thomas Vector Fields
Author(s): Oh, Duk-Soon
Abstract:
Raviart-Thomas finite elements are very useful for problems posed in H(div) since they are H(div)-conforming. We introduce two domain decomposition methods for solving vector field problems posed in H(div) discretized by Raviart-Thomas finite elements.
A two-level overlapping Schwarz method is developed. The coarse part of the preconditioner is based on energy-minimizing extensions and the local parts consist of traditional solvers on overlapping subdomains. We prove that our method is scalable and that the condition number grows linearly with the logarithm of the number of degrees of freedom in the individual subdomains and linearly with the relative overlap between the overlapping subdomains. The condition number of the method is also independent of the values and jumps of the coefficients across the interface between subdomains. We provide numerical results to support our theory.
We also consider a balancing domain decomposition by constraints (BDDC) method. The BDDC preconditioner consists of a coarse part involving primal constraints across the interface between subdomains and local parts related to the Schur complements corresponding to the local subdomain problems. We provide bounds of the condition number of the preconditioned linear system and suggest that the condition number has a polylogarithmic bound in terms of the number of degrees of freedom in the individual subdomains from our numerical experiments for arbitrary jumps of the coefficients across the subdomain interfaces.
Title: From a Calculus to an Execution Environment for Stream Processing
Author(s): Soulé, Robert; Hirzel, Martin; Gedik, Bugra; Grimm, Robert
Abstract:
At one level, this paper is about River, a virtual execution environment for stream processing. Stream processing is a paradigm well-suited for many modern data processing systems that ingest high-volume data streams from the real world, such as audio/video streaming, high-frequency trading, and security monitoring. One attractive property of stream processing is that it lends itself to parallelization on multi-cores, and even to distribution on clusters when extreme scale is required. Stream processing has been coevolved by several communities, leading to diverse languages with similar core concepts. Providing a common execution environment reduces language development effort and increases portability. We designed River as a practical realization of Brooklet, a calculus for stream processing. So at another level, this paper is about a journey from theory (the calculus) to practice (the execution environment). The challenge is that, by definition, a calculus abstracts away all but the most central concepts. Hence, there are several research questions in concretizing the missing parts, not to mention a significant engineering effort in implementing them. But the effort is well worth it, because the benefit of using a calculus as a foundation is that it yields clear semantics and proven correctness results.
Title: Cryptographic Resilience to Continual Information Leakage
Candidate: Wichs, Daniel
Advisor(s): Dodis, Yevgeniy
Abstract:
We study the question of achieving cryptographic security on devices that leak information about their internal secret state to an external attacker.This study is motivated by the prevalence of side-channel attacks, where the physical characteristics of a computation (e.g. timing, power-consumption, temperature, radiation, acoustics, etc.) can be measured, and may reveal useful information about the internal state of a device. Since some such leakage is inevitably present in almost any physical implementation, we believe that this problem cannot just be addressed by physical countermeasures alone. Instead, it should already be taken into account when designing the mathematical specification of cryptographic primitives and included in the formal study of their security.
In this thesis, we propose a new formal framework for modeling the leakage available to an attacker. This framework, called the continual leakage model, assumes that an attacker can continually learn arbitrary information about the internal secret state of a cryptographic scheme at any point in time, subject only to the constraint that the rate of leakage is bounded. More precisely, our model assumes some abstract notion of time periods. In each such period, the attacker can choose to learn arbitrary functions of the current secret state of the scheme, as long as the number of output bits leaked is not too large. In our solutions, cryptographic schemes will continually update their internal secret state at the end of each time period. This will ensure that leakage observed in different time periods cannot be meaningfully combined to break the security of the cryptosystem. Although these updates modify the secret state of the cryptosystem, the desired functionality of the scheme is preserved, and the users can remain oblivious to these updates. We construct signatures, encryption, and secret sharing/storage schemes in this model.
Title: Surface Representation of Particle Based Fluids
Candidate: Yu, Jihun
Advisor(s): Yap, Chee
Abstract:
In this thesis, we focus on surface representation for particle-based fluid simulators such as Smoothed Particle Hydrodynamics (SPH). We first present a new surface reconstruction algorithm which formulates the implicit function as a sum of anisotropic smoothing kernels. The direction of anisotropy at a particle is determined by performing Weighted Principal Component Analysis (WPCA) over the neighboring particles. In addition, we perform a smoothing step that re-positions the centers of these smoothing kernels. Since these anisotropic moothing kernels capture the local particle distributions more accurately, our method has advantages over existing methods in representing smooth surfaces, thin streams and sharp features of fluids. This method is fast, easy to implement, and the results demonstrate a significant improvement in the quality of reconstructed surfaces as compared to existing methods. Next,we introduce the idea of using an explicit triangle mesh to track the air/liquid interface in a SPH simulator.
Once an initial surface mesh is created, this mesh is carried forward in time using nearby particle velocities to advect the mesh vertices. The mesh connectivity remains mostly unchanged across time-steps; it is only modified locally for topology change events or for the improvement of triangle quality. In order to ensure that the surface mesh does not diverge from the underlying particle simulation, we periodically project the mesh surface onto an implicit surface defined by the physics simulation. The mesh surface presents several advantages over previous SPH surface tracking techniques: A new method for surface tension calculations clearly outperforms the state of the art in SPH surface tension for computer graphics. A new method for tracking detailed surface information (like colors) is less susceptible to numerical diffusion than competing techniques. Finally, a temporally-coherent surface mesh allows us to simulate high-resolution surface wave dynamics without being limited by the particle resolution of the SPH simulation.
Title: Design and Results of the 4th Annual Satisfiability Modulo Theories Competition (SMT-COMP 2008)
Author(s): Barrett, Clark; Deters, Morgan; Oliveras, Albert; Stump, Aaron
Abstract:
The Satisfiability Modulo Theories Competition (SMT-COMP) is an annual competition aimed at stimulating the advance of the state-of-the-art techniques and tools developed by the Satisfiability Modulo Theories (SMT) community. As with the first three editions, SMT-COMP 2008 was held as a satellite event of CAV 2008, held July 7-14, 2008. This report gives an overview of the rules, competition format, benchmarks, participants and results of SMT-COMP 2008.
Title: DTAC: A method for planning to claim in Bridge
Candidate: Bethe, Paul
Advisor(s): Davis, Ernest
Abstract:
The DTAC program uses depth-first search to find an unconditional claim in bridge; that is, a line of play that is guaranteed to succeed whatever the distribution of the outstanding cards among the defenders. It can also find claims that are guaranteed to succeed under specified assumptions about the distribution of the defenders. cards. Lastly, DTAC can find a claim which requires losing a trick at some point. Using transposition tables to detect repeated positions, DTAC can carry out a complete DFS to find an unconditional ordered claim in less than 0.001 seconds on average, and less than 1 second for claims which lose a trick. The source code for DTAC is available from: http://cs.nyu.edu/~pmb309/DTAC.html
Title: On the Randomness Requirements for Privacy
Candidate: Bosley, Carleton
Advisor(s): Dodis, Yevgeniy
Abstract:
Most cryptographic primitives require randomness (for example, to generate secret keys). Usually, one assumes that perfect randomness is available, but, conceivably, such primitives might be built under weaker, more realistic assumptions. This is known to be achievable for many authentication applications, when entropy alone is typically sufficient. In contrast, all known techniques for achieving privacy seem to fundamentally require (nearly) perfect randomness. We ask the question whether this is just a coincidence, or, perhaps, privacy inherently requires true randomness?
We completely resolve this question for information-theoretic private-key encryption, where parties wish to encrypt a b-bit value using a shared secret key sampled from some imperfect source of randomness S. Our technique also extends to related primitives which are sufficiently binding and hiding, including computationally secure commitments and public-key encryption.
Our main result shows that if such n-bit source S allows for a secure encryption of b bits, where b > log n, then one can deterministically extract nearly b almost perfect random bits from S . Further, the restriction that b > log n is nearly tight: there exist sources S allowing one to perfectly encrypt (log n - log log n) bits, but not to deterministically extract even a single slightly unbiased bit.
Hence, to a large extent, true randomness is inherent for encryption: either the key length must be exponential in the message length b, or one can deterministically extract nearly b almost unbiased random bits from the key. In particular, the one-time pad scheme is essentially "universal".
Title: Machine Learning Approaches to Gene Duplication and Transcription Regulation
Candidate: Chen, Huang-Wen
Advisor(s): Shasha, Dennis
Abstract:
Gene duplication can lead to genetic redundancy or functional divergence, when duplicated genes evolve independently or partition the original function. In this dissertation, we employed machine learning approaches to study two different views of this problem: 1) Redundome, which explored the redundancy of gene pairs in the genome of Arabidopsis thaliana, and 2) ContactBind, which focused on functional divergence of transcription factors by mutating contact residues to change binding affinity.
In the Redundome project, we used machine learning techniques to classify gene family members into redundant and non-redundant gene pairs in Arabidopsis thaliana, where sufficient genetic and genomic data is available. We showed that Support Vector Machines were two-fold more precise than single attribute classifiers, and performed among the best within other machine learning algorithms. Machine learning methods predict that about half of all genes in Arabidopsis showed the signature of predicted redundancy with at least one but typically less than three other family members. Interestingly, a large proportion of predicted redundant gene pairs were relatively old duplications (e.g., Ks>1), suggesting that redundancy is stable over long evolutionary periods. The genome-wide predictions were plot with similarity trees based on ClustalW alignment scores, and can be accessed at http://redundome.bio.nyu.edu .
In the ContactBind project, we use Bayesian networks to model dependences between contact residues in transcription factors and binding site sequences. Based on the models learned from various binding experiments, we predicted binding motifs and their locations on promoters for three families of transcription factors in three species. The predictions are publicly available at http://contactbind.bio.nyu.edu . The website also provides tools to predict binding motifs and their locations for novel protein sequences of transcription factors. Users can construct their Bayesian networks for new families once such a familial binding data is available.
Title: New Privacy-Preserving Architectures for Identity-/Attribute-based Encryption
Candidate: Chow, Sze Ming
Advisor(s): Dodis, Yevgeniy; Shoup, Victor
Abstract:
The notion of identity-based encryption (IBE) was proposed as an economical alternative to public-key infrastructures. IBE is also a useful building block in various cryptographic primitives such as searchable encryption. A generalization of IBE is attribute-based encryption (ABE). A major application of ABE is fine-grained cryptographic access control of data. Research on these topics is still actively continuing.
However, security and privacy of IBE and ABE are hinged on the assumption that the authority which setups the system is honest. Our study aims to reduce this trust assumption.
The inherent key escrow of IBE has sparkled numerous debates in the cryptography/security community. A curious key generation center (KGC) can simply generate the user's private key to decrypt a ciphertext. However, can a KGC still decrypt if it does not know the intended recipient of the ciphertext? This question is answered by formalizing KGC anonymous ciphertext indistinguishability (ACI-KGC). All existing practical pairing-based IBE schemes without random oracles do not achieve this notion. In this thesis, we propose an IBE scheme with ACI-KGC, and a new system architecture with an anonymous secret key generation protocol such that the KGC can issue keys to authenticated users without knowing the list of users' identities. This also matches the practice that authentication should be done with the local registration authorities. Our proposal can be viewed as mitigating the key escrow problem in a new dimension.
For ABE, it is not realistic to trust a single authority to monitor all attributes and hence distributing control over many attribute-authorities is desirable. A multi-authority ABE scheme can be realized with a trusted central authority (CA) which issues part of the decryption key according to a user's global identifier (GID). However, this CA may have the power to decrypt every ciphertext, and the use of a consistent GID allowed the attribute-authorities to collectively build a full profile with all of a user's attributes. This thesis proposes a solution without the trusted CA and without compromising users' privacy, thus making ABE more usable in practice.
Underlying both contributions are our new privacy-preserving architectures enabled by borrowing techniques from anonymous credential.
Title: Coordination Mechanisms for Weighted Sum of Completion Times
Author(s): Cole, Richard; Gkatzelis, Vasilis; Mirrokni, Vahab
Abstract:
We study policies aiming to minimize the weighted sum of completion times of jobs in the context of coordination mechanisms for selfish scheduling problems. Our goal is to design local policies that achieve a good price of anarchy in the resulting equilibria for unrelated machine scheduling. In short, we present the first constant-factor-approximate coordination mechanisms for this model.
First, we present a generalization of the ShortestFirst policy for weighted jobs, called SmithRule; we prove that it achieves an approximation ratio of 4 and we show that any set of non-preemptive ordering policies can result in equilibria with approximation ratio at least 3 even for unweighted jobs. Then, we present ProportionalSharing, a preemptive strongly local policy that beats this lower bound of 3; we show that this policy achieves an approximation ratio of 2.61 for the weighted sum of completion times and that the EqualSharing policy achieves an approximation ratio of 2.5 for the (unweighted) sum of completion times. Furthermore, we show that ProportionalSharing induces potential games (in which best-response dynamics converge to pure Nash equilibria).
All of our upper bounds are for the robust price of anarchy, defined by Roughgarden [36], so they naturally extend to mixed Nash equilibria, correlated equilibria, and regret minimization dynamics. Finally, we prove that our price of anarchy bound for ProportionalSharing can be used to design a new combinatorial constant-factor approximation algorithm minimizing weighted completion time for unrelated machine scheduling.
Title: Tools and Techniques for the Sound Verification of Low Level Code
Candidate: Conway, Christopher L.
Advisor(s): Barrett, Clark
Abstract:
Software plays an increasingly crucial role in nearly every facet of modern life, from communications infrastructure to the control systems in automobiles, airplanes, and power plants. To achieve the highest degree of reliability for the most critical pieces of software, it is necessary to move beyond ad hoc testing and review processes towards verification---to prove using formal methods that a piece of code exhibits exactly those behaviors allowed by its specification and no others.
A significant portion of the existing software infrastructure is written in low-level languages like C and C++. Features of these language present significant verification challenges. For example, unrestricted pointer manipulation means that we cannot prove even the simplest properties of programs without first collecting precise information about potential aliasing relationships between variables.
In this thesis, I present several contributions. The first is a general framework for combining program analyses that are only conditionally sound. Using this framework, I show it is possible to design a sound verification tool that relies on a separate, previously-computed pointer analysis.
The second contribution of this thesis is Cascade, a multi-platform, multi-paradigm framework for verification. Cascade includes a support for precise analysis of low-level C code, as well as for higher-level languages such as SPL.
Finally, I describe a novel technique for the verification of datatype invariants in low-level systems code. The programmer provides a high-level specification for a low-level implementation in the form of inductive datatype declarations and code assertions. The connection between the high-level semantics and the implementation code is then checked using bit-precise reasoning. An implementation of this datatype verification technique is available as a Cascade module.
Title: Probabilistic and Topological methods in Computational Geometry
Candidate: Dhandapani, Raghavan
Advisor(s): Pach, Janos
Abstract:
We consider four problems connected by the common thread of geometry. The first three involve problems and algorithms that arise in applications that apriori do not involve geometry but this turns out to be the right language for visualizing and analyzing them. In the fourth, we generalize some well known results in geometry to the topological plane. The techniques we use come from probability and topology.
First, we consider two algorithms that work well in practice but the theoretical mechanism behind whose success is not very well understood.
Greedy routing is a routing mechanism that is commonly used in wireless sensor networks. While routing on the Internet uses standard established protocols, routing in ad-hoc networks with little structure (like sensor networks) is more difficult. Practitioners have devised algorithms that work well in practice, however they were no known theoretical guarantees. We provide the first such result in this area by showing that greedy routing can be made to work on Planar triangulations.
Linear Programming is a technique for optimizing a linear function subject to linear constraints. Simplex Algorithms are a family of algorithms that have proven quite successful in solving Linear Programs in practice. However, examples of Linear Programs on which these algorithms are very inefficient have been obtained by researchers. In order to explain this discrepancy between theory and practice, many authors have shown that Simplex Algorithms are efficient in expectation on randomized Linear Programs. We strengthen these results by proving a partial concentration bound for the Shadow Vertex Simplex Algorithm.
Next, we point out a limitation in an algorithm that is used commonly by practitioners and suggest a way of overcoming this.
Recommendation Systems are algorithms that are used to recommend goods (books, movies etc.) to users based on the similarities between their past preferences and those of other users. Low Rank Approximation is a common method used for this. We point out a common limitation of this method in the presence of ill-conditioning: the presence of multiple local minima. We also suggest a simple averaging based technique to overcome this limitation.
Finally, we consider some basic results in convexity like Radon's, Helly's and Caratheodory's theorems and generalize them to the topological plane, i.e., a plane which has the concept of a linear path which is analogous to a straight line but no notion of metric or distances.
Title: An Iterative Substructuring Algorithm for Two-dimensional Problems in H(curl)
Author(s): Dohrmann, Clark R.; Widlund, Olof B.
Abstract:
A domain decomposition algorithm, similar to classical iterative substructuring algorithms, is presented for two-dimensional problems in the space H0(curl). It is defined in terms of a coarse space and local subspaces associated with individual edges of the subdomains into which the domain of the problem has been subdivided. The algorithm differs from others in three basic respects. First, it can be implemented in an algebraic manner that does not require access to individual subdomain matrices or a coarse discretization of the domain; this is in contrast to algorithms of the BDDC, FETIâDP, and classical twoâlevel overlapping Schwarz families. Second, favorable condition number bounds can be established over a broader range of subdomain material properties than in previous studies. Third, we are able to develop theory for quite irregular subdomains and bounds for the condition number of our preconditioned conjugate gradient algorithm, which depend only on a few geometric parameters.
The coarse space for the algorithm is based on simple energy minimization concepts, and its dimension equals the number of subdomain edges. Numerical results are presented which confirm the theory and demonstrate the usefulness of the algorithm for a variety of mesh decompositions and distributions of material properties.
Title: Semi-Supervised Learning via Generalized Maximum Entropy
Candidate: Erkan, Ayse Naz
Advisor(s): LeCun, Yann
Abstract:
Maximum entropy (MaxEnt) framework has been studied extensively in the supervised setting. Here, the goal is to find a distribution p, that maximizes an entropy function while enforcing data constraints so that the expected values of some (pre-defined) features with respect to p, match their empirical counterparts approximately. Using different entropy measures, different model spaces for p and different approximation criteria for the data constraints yields a family of discriminative supervised learning methods (e.g., logistic regression, conditional random fields, least squares and boosting). This framework is known as the generalized maximum entropy framework.
Semi-supervised learning (SSL) has emerged in the last decade as a promising field that enables utilizing unlabeled data along with labeled data so as to increase the accuracy and robustness of inference algorithms. However, most SSL algorithms to date have had trade-offs, for instance in terms of scalability or applicability to multi-categorical data.
In this thesis, we extend the generalized MaxEnt framework to develop a family of novel SSL algorithms using two different approaches: i. Introducing Similarity Constraints We incorporate unlabeled data via modifications to the primal MaxEnt objective in terms of additional potential functions. A potential function stands for a closed proper convex function that can take the form of a constraint and/or a penalty representing our structural assumptions on the data geometry. Specifically, we impose similarity constraints as additional penalties based on the semi-supervised smoothness assumption; i.e., we restrict the generalized MaxEnt problem such that similar samples have similar model outputs. ii. Augmenting Constraints on Model Features We incorporate unlabeled data to enhance the estimates on the model and empirical expectations based on our assumptions on the data geometry.
In particular, we derive the semi-supervised formulations for three specific instances of the generalized MaxEnt on conditional distributions, namely logistic regression and kernel logistic regression for multi-class problems, and conditional random fields for structured output prediction problems. A thorough empirical evaluation on standard data sets that are widely used in the literature demonstrates the validity and competitiveness of the proposed algorithms. In addition to these benchmark data sets, we apply our approach to two real-life problems: i. vision based robot grasping, and ii. remote sensing image classification, where the scarcity of the labeled training samples is the main bottleneck in the learning process. For the particular case of grasp learning, we propose a combination of semi-supervised learning and active learning, another sub-field of machine learning that is focused on the scarcity of labeled samples, when the problem setup is suitable for incremental labeling.
The novel SSL algorithms proposed in this thesis have numerous advantages over the existing semi-supervised algorithms as they yield convex, scalable, inherently multi-class loss functions that can be kernelized naturally.
Title: Information Extraction on High-School Level Chemistry Labs
Author(s): Galron, Daniel
Abstract:
In this report we present a feasibility study on automatically interpreting instructions found in a set of high school chemistry labs, and discuss the role of deep domain knowledge in the interpretation. We define the task of sentence-level interpretation as the extraction of symbolic representations of the sentence semantics. In the broader scope, the sentence-level semantics of a particular sentence will be resolved with semantics from other sentences in the lab along with domain knowledge to disambiguate and reason about a physical system. The task of general automatic sentence-level interpretation is a difficult one. The general problem is not very well defined in the natural language processing research community, and few researchers have studied the problem. The common practice is to decompose the problem into subtasks, such as resolving coreferences of noun phrases, labeling the semantic roles of arguments to predicates, and identifying word categories. We describe a pipeline combining the subtasks described, along with parsing, to create a system capable of extracting sentence-level semantics. All the systems used for the subtask are found off-the-shelf, and we should stress that such a system will be highly-error prone for reasons we discuss. Finally, we do a close study of the chemistry lab corpus, and analyze each instruction to determine the feasibility of its automatic interpretation and the role of deep domain knowledge in its disambiguation and understanding.
Title: Solving Quantified First Order Formulas in Satisfiability Modulo Theories
Candidate: Ge, Yeting
Advisor(s): Barrett, Clark
Abstract:
Design errors in computer systems, i.e. bugs, can cause inconvenience, loss of data and time, and in some cases catastrophic damages. One approach for improving design correctness is formal methods: techniques aiming at mathematically establishing that a piece of hardware or software satisfies certain properties. For some industrial cases in which formal methods are utilized, quantified first order formulas in satisfiability modulo theories (SMT) are useful. This dissertation presents several novel techniques for solving quantified formulas in SMT.
In general, deciding a quantified formula in SMT is undecidable. The practical approach for general quantifier reasoning in SMT is heuristics-based instantiation. This dissertation proposes a number of new heuristics that solves several challenges. Experimental results show that with the new heuristics a significant number of more benchmarks can be solved than before.
When only consider formulas within certain fragments of first order logic, it is possible to have complete algorithms based on instantiation. We propose several new fragments, and we prove that formulas in these fragments can be solved by a complete algorithm based on instantiation. For satisfiable quantified formulas in these fragments, we show how to construct the models.
As SMT solvers grow in complexity, the correctness of SMT solvers become questionable. A practical method to improve the correctness is to check the proofs from SMT solvers. We propose a proof translator that translates proofs from SMT solver CVC3 into a trusted solver HOL Light that actually checks the proofs. Experiments with the proof translator discover a faulty proof rule in CVC3 and two MIT-labeled quantified benchmarks in the SMT benchmark library SMT-LIB.
Title: Polite Theories Revisited
Author(s): Jovanovic, Dejan; Barrett, Clark
Abstract:
The classic method of Nelson and Oppen for combining decision procedures requires the theories to be stably-infnite. Unfortunately, some important theories do not fall into this category (e.g. the theory of bit-vectors). To remedy this problem, previous work introduced the notion of polite theories. Polite theories can be combined with any other theory using an extension of the Nelson-Oppen approach. In this paper we revisit the notion of polite theories, fxing a subtle flaw in the original definition. We give a new combination theorem which specifies the degree to which politeness is preserved when combining polite theories. We also give conditions under which politeness is preserved when instantiating theories by identifying two sorts. These results lead to a more general variant of the theorem for combining multiple polite theories.
Title: TestRig: A Platform independent system testing tool
Candidate: Kaul, Vaibhav
Advisor(s): Shasha, Dennis
Abstract:
The goal of the TestRig software is to give a test engineer a fixed interface to help him with system/integration testing of software systems. TestRig is platform independent and can be utilized to test software systems coded with any programming language. In addition to doing that, it provides templates and examples of using various Open Source testing tools to help a user design their test cases. TestRig has been designed keeping in mind the current scenario in software development where complex systems are often created using multiple programming languages across different platforms. The challenge is to have a defined set of rules that are able to test any such system. The software makes use of various open source testing tools to run tests and verify results, which enables a user to test a system at different levels such as Performance Testing, Blackbox Testing, and User Acceptance Testing. TestRig is open source and utilizes a programmer’s creativity to test across multiple scenarios. The thesis will show how different software systems have been tested using TestRig.
Title: An Algorithmic Enquiry Concerning Causality
Candidate: Kleinberg, Samantha
Advisor(s): Mishra, Bhubaneswar
Abstract:
In many domains we face the problem of determining the underlying causal structure from time-course observations of a system. Whether we have neural spike trains in neuroscience, gene expression levels in systems biology, or stock price movements in finance, we want to determine why these systems behave the way they do. For this purpose we must assess which of the myriad possible causes are significant while aiming to do so with a feasible computational complexity. At the same time, there has been much work in philosophy on what it means for something to be a cause, but comparatively little attention has been paid to how we can identify these causes. Algorithmic approaches from computer science have provided the first steps in this direction, but fail to capture the complex, probabilistic and temporal nature of the relationships we seek.
This dissertation presents a novel approach to the inference of general (type-level) and singular (token-level) causes. The approach combines philosophical notions of causality with algorithmic approaches built on model checking and statistical techniques for false discovery rate control. By using a probabilistic computation tree logic to describe both cause and effect, we allow for complex relationships and explicit description of the time between cause and effect as well as the probability of this relationship being observed (e.g. "a and b until c, causing d in 10-20 time units"). Using these causal formulas and their associated probabilities, we develop a novel measure for the significance of a cause for its effect, thus allowing discovery of those that are statistically interesting, determined using the concepts of multiple hypothesis testing and false discovery control. We develop algorithms for testing these properties in time-series observations and for relating the inferred general relationships to token-level events (described as sequences of observations). Finally, we illustrate these ideas with example data from both neuroscience and finance, comparing the results to those found with other inference methods. The results demonstrate that our approach achieves superior control of false discovery rates, due to its ability to appropriately represent and infer temporal information.
Title: The Temporal Logic of Token Causes
Author(s): Kleinberg, Samantha; Mishra, Bud
Abstract:
While type causality helps understand general relationships such as the etiology of a disease (smoking causing lung cancer), token causality aims to explain causal connections in specific instantiated events, such as the diagnosis of a specific patient (Ravi's developing lung cancer after a 20-year smoking habit). Understanding why something happened, as in these examples, is central to reasoning in such diverse cases as the diagnosis of patients, understanding why the US financial market collapsed in 2007 and finding a causal explanation for Obama's victory over Clinton in the US primary. However, despite centuries of work in philosophy and decades of research in computer science, the problem of how to rigorously formalize token causality and how to automate such reasoning has remained unsolved. In this paper, we show how to use type-level causal relationships, represented as temporal logic formulas, together with philosophical principles, to reason about these token-level cases. Finally, we show how this method can correctly reason about examples that have traditionally proven difficult for both computational and philosophical theories to handle.
Title: An overlapping domain decomposition method for the Reissner-Mindlin Plate with the Falk-Tu Elements
Author(s): Lee, Jong Ho
Abstract:
The Reissner-Mindlin plate theory models a thin plate with thickness t. The condition numbers of finite element approximations of this model deteriorate badly as the thickness t of the plate converges to 0. In this paper, we develop an overlapping domain decomposition method for the Reissner-Mindlin plate model discretized by the Falk-Tu elements with the convergence rate which does not deteriorate when t converges to 0. It is shown that the condition number of this overlapping method is bounded by C(1+ H/delta)^3(1 +logH/h)^2. Here H is the maximum diameter of the subdomains, delta the size of overlap between subdomains, and h the element size. Numerical examples are provided to confirm the theory.
Title: Time Series Modeling with Hidden Variables and Gradient-Based Algorithms
Candidate: Mirowski, Piotr
Advisor(s): LeCun, Yann
Abstract:
We collect time series from real-world phenomena, such as gene interactions in biology or word frequencies in consecutive news articles. However, these data present us with an incomplete picture, as they result from complex dynamical processes involving unobserved state variables. Research on state-space models is motivated by simultaneously trying to infer hidden state variables from observations, as well as learning the associated dynamic and generative models.
I have developed a tractable, gradient-based method for training Dynamic Factor Graphs (DFG) with continuous latent variables. A DFG consists of (potentially nonlinear) factors modeling joint probabilities between hidden and observed variables. The DFG assigns a scalar energy to each configuration of variables, and a gradient-based inference procedure finds the minimum-energy state sequence for a given observation sequence. We approximate maximum likelihood learning by minimizing the expected energy over training sequences with respect to the factors' parameters. These alternated inference and parameter updates constitute a deterministic EM-like procedure.
Using nonlinear factors such as deep, convolutional networks, DFGs were shown to reconstruct chaotic attractors, to outperform a time series prediction benchmark, and to successfully impute motion capture data where a large number of markers were missing. In a joint work with the NYU Plant Systems Biology Lab, DFGs have been subsequently employed to the discovery of gene regulation networks by learning the dynamics of mRNA expression levels.
DFGs have also been extended into a deep auto-encoder architecture, and used on time-stamped text documents, with word frequencies as inputs. We focused on collections of documents that exhibit a structure over time. Working as dynamic topic models, DFGs could extract a latent trajectory from consecutive political speeches; applied to news articles, they achieved state-of-the-art text categorization and retrieval performance.
Finally, I used an embodiment of DFGs to evaluate the likelihood of discrete sequences of words in text corpora, relying on dynamics on word embeddings. Collaborating with AT&T; Labs Research on a project in speech recognition, we have improved on existing continuous statistical language models by enriching them with word features and long-range topic dependencies.
Title: An Overlapping Schwarz Algorithm for Raviart-Thomas Vector Fields with Discontinuous Coefficients
Author(s): Oh, Duk-Soon
Abstract:
Overlapping Schwarz methods form one of two major families of domain decomposition methods. We consider a two-level overlapping Schwarz method for Raviart-Thomas vector fields. The coarse part of the preconditioner is based on the energy-minimizing extensions and the local parts are based on traditional solvers on overlapping subdomains. We show that the condition number grows linearly with the logarithm of the number of degrees of freedom in the individual subdomains and linearly with the relative overlap between the overlapping subdomains. The condition number of the method is also independent of the values and jumps of the coefficients. Numerical results for 2D and 3D problems, which support the theory, are also presented.
Title: BDDC preconditioners for spectral element discretizations of almost incompressible elasticity in three dimensions
Author(s): Pavarino, Luca F.; Widlund, Olof B.; Zampini, Stefano
Abstract:
BDDC algorithms are constructed and analyzed for the system of almost incompressible elasticity discretized with Gauss-Lobatto-Legendre spectral elements in three dimensions. Initially mixed spectral elements are employed to discretize the almost incompressible elasticity system, but a positive definite reformulation is obtained by eliminating all pressure degrees of freedom interior to each subdomain into which the spectral elements have been grouped. Appropriate sets of primal constraints can be associated with the subdomain vertices, edges, and faces so that the resulting BDDC methods have a fast convergence rate independent of the almost incompressibility of the material. In particular, the condition number of the BDDC preconditioned operator is shown to depend only weakly on the polynomial degree $n$, the ratio $H/h$ of subdomain and element diameters, and the inverse of the inf-sup constants of the subdomains and the underlying mixed formulation, while being scalable, i.e., independent of the number of subdomains and robust, i.e., independent of the Poisson ratio and Young's modulus of the material considered. These results also apply to the related FETI-DP algorithms defined by the same set of primal constraints. Numerical experiments carried out on parallel computing systems confirm these results.
Title: Structure Prediction and Visualization in Molecular Biology
Candidate: Poultney, Christopher
Advisor(s): Shasha, Dennis
Abstract:
The tools of computer science can be a tremendous help to the working biologist. Two broad areas where this is particularly true are visualization and prediction. In visualization, the size of the data involved often makes meaningful exploration of the data and discovery of salient features difficult and time-consuming. Similarly, intelligent prediction algorithms can greatly reduce the lab time required to achieve significant results, or can reduce an intractable space of potential experiments to a tractable size.
Whereas the thesis discusses both a visualization technique and a machine learning problem, the thesis presentation will focus exclusively on the machine learning problem: prediction of temperature-sensitive mutations from protein structure. Temperature-sensitive mutations are a tremendously valuable research tool particularly for studying genes such as yeast essentially genes. To date, most methods for generating temperature-sensitive mutations involve large-scale random mutations followed by an intensive screening and characterization process. While there have been successful efforts to improve this process by rational design of temperature-sensitive proteins, surprisingly little work has been done in the area of predicting those mutations that will exhibit a temperature-sensitive phenotype. We describe a system that, given the structure of a protein of interest, uses a combination of protein structure prediction and machine learning to provide a ranked "top 5" list of likely candidates for temperature-sensitive mutations.
Title: An Empirical Bayesian Interpretation and Generalization of NL-means
Author(s): Raphan, Martin; Simoncelli, Eero P.
Abstract:
A number of recent algorithms in signal and image processing are based on the empirical distribution of localized patches. Here, we develop a nonparametric empirical Bayesian estimator for recovering an image corrupted by additive Gaussian noise, based on fitting the density over image patches with a local exponential model. The resulting solution is in the form of an adaptively weighted average of the observed patch with the mean of a set of similar patches, and thus both justifies and generalizes the recently proposed nonlocal-means (NL-means) method for image denoising. Unlike NL-means, our estimator includes a dependency on the size of the patch similarity neighborhood, and we show that this neighborhood size can be chosen in such a way that the estimator converges to the optimal Bayes least squares estimator as the amount of data grows. We demonstrate the increase in performance of our method compared to NL-means on a set of simulated examples.
Title: Theoretical Foundations and Algorithms for Learning with Multiple Kernels
Candidate: Rostamizadeh, Afshin
Advisor(s): Mohri, Mehryar
Abstract:
Kernel-based algorithms have been used with great success in a variety of machine learning applications. These include algorithms such as support vector machines for classification, kernel ridge regression, ranking algorithms, clustering algorithms, and virtually all popular dimensionality reduction algorithms, since they are special instances of kernel principal component analysis.
But, the choice of the kernel, which is crucial to the success of these algorithms, has been traditionally left entirely to the user. Rather than requesting the user to commit to a specific kernel, multiple kernel algorithms require the user only to specify a family of kernels. This family of kernels can be used by a learning algorithm to form a combined kernel and derive an accurate predictor. This is a problem that has attracted a lot of attention recently, both from the theoretical point of view and from the algorithmic, optimization, and application point of view.
This thesis presents a number of novel theoretical and algorithmic results for learning with multiple kernels.
It gives the first tight margin-based generalization bounds for learning kernels with Lp regularization. In particular, our margin bounds for L1 regularization are shown to have only a logarithmic dependency on the number of kernels, which is a significant improvement over all previous analyses. Our results also include stability-based guarantees for a class of regression algorithms. In all cases, these guarantees indicate the benefits of learning with a large number of kernels.
We also present a family of new two-stage algorithms for learning kernels based on a notion of alignment and give an extensive analysis of the properties of these algorithms. We show the existence of good predictors for the notion of alignment we define and give efficient algorithms for learning a maximum alignment kernel by showing that the problem can be reduced to a simple QP.
Finally, we also report the results of extensive experiments with our two-stage algorithms in classification and regression tasks, which show an improvement both over the uniform combination of kernels and over other state-of-the-art learning kernel methods for L1 and L2 regularization. These might constitute the first series of results for learning with multiple kernels that demonstrate a consistent improvement over a uniform combination of kernels.
Title: Creating collections and evaluating viewpoints: Selection techniques for interface design
Candidate: Secord, Adrian
Advisor(s): Zorin, Denis
Abstract:
In computer graphics and user interface design, selection problems are those that require the user to select a collection consisting of a small number of items from a much larger library. This dissertation explores selection problems in two diverse domains: large personal multimedia collections, containing items such as personal photographs or songs, and camera positions for 3D objects, where each item is a different viewpoint observing an object. Multimedia collections have by discrete items with strong associated metadata, while camera positions form a continuous space but are weak in metadata. In either domain, the items to be selected have rich interconnections and dependencies, making it difficult to successfully apply simple techniques (such as ranking) to aid the user. Accordingly, we develop separate approaches for the two domains.
For personal multimedia collections, we leverage the semantic metadata associated with each item (such as song title, artist name, etc.) and provide the user with a simple query language to describe their desired collection. Our system automatically suggests a collection of items that conform to the userâs query. Since any query language has limited expressive power, and since users often create collections via exploration, we provide various refinement techniques that allow the user to expand, refine and explore their collection directly through examples.
For camera positioning, we do not have the advantage of having semantic metadata for each item, unlike in media collections. We instead create a proxy viewpoint goodness function which can be used to guide the solution of various selection problems involving camera viewpoints. This function is constructed from several different attributes of the viewpoint, such as how much surface area is visible, or how "curvy" the silhouette is. Since there are many possible viewpoint goodness functions, we conducted a large user study of viewpoint preference and use the results to evaluate thousands of different functions and find the best ones. While we suggest several goodness functions to the practitioner, our user study data and methodology can be used to evaluate any proposed goodness function; we hope it will be a useful tool for other researchers.
Title: Henrique Andrade, Vibhore Kumar, and Kun-Lung Wu, A Universal Calculus for Stream Processing Languages
Author(s): Soulé, Robert; Hirzel, Martin; Grimm, Robert; Gedik, Buğra
Abstract:
Stream processing applications such as algorithmic trading, MPEG processing, and web content analysis are ubiquitous and essential to business and entertainment. Language designers have developed numerous domain-specific languages that are both tailored to the needs of their applications, and optimized for performance on their particular target platforms. Unfortunately, the goals of generality and performance are frequently at odds, and prior work on the formal semantics of stream processing languages does not capture the details necessary for reasoning about implementations. This paper presents Brooklet, a core calculus for stream processing that allows us to reason about how to map languages to platforms and how to optimize stream programs. We translate from three representative languages, CQL, StreamIt, and Sawzall, to Brooklet, and show that the translations are correct. We formalize three popular and vital optimizations, data-parallel computation, operator fusion, and operator re-ordering, and show under which conditions they are correct. Language designers can use Brooklet to specify exactly how new features or languages behave. Language implementors can use Brooklet to show exactly under which circumstances new optimizations are correct. In ongoing work, we are developing an intermediate language for streaming that is based on Brooklet. We are implementing our intermediate language on System S, IBM's high-performance streaming middleware.
Title: Analysis of Mass Spectrometry Data for Protein Identification In Complex Biological Mixtures
Candidate: Spivak, Marina
Advisor(s): Greengard, Leslie
Abstract:
Mass spectrometry is a powerful technique in analytical chemistry that was originally designed to determine the composition of small molecules in terms of their constituent elements. In the last several decades, it has begun to be used for much more complex tasks, including the detailed analysis of the amino acid sequence that makes up an unknown protein and even the identification of multiple proteins present in a complex mixture. The latter problem is largely unsolved and the principal subject of this dissertation.
The fundamental difficulty in the analysis of mass spectrometry data is that of ill-posedness. There are multiple solutions consistent with the experimental data and the data is subject to significant amounts of noise. In this work, we have developed application-specific machine learning algorithms that (partially) overcome this ill-posedness. We make use of labeled examples of a single class of peptide fragments and of the unlabeled fragments detected by the instrument. This places the approach within the broader framework of semi-supervised learning.
Recently, there has been considerable interest in classification problems of this type, where the learning algorithm only has access to labeled examples of a single class and unlabeled data. The motivation for such problems is that in many applications, examples of one of the two classes are easy and inexpensive to obtain, whereas the acquisition of examples of a second class is difficult and labor-intensive. For example, in document classification, positive examples are documents that address specific subject, while unlabeled documents are abundant. In movie rating, the positive data are the movies chosen by clients, while the unlabeled data are all remaining movies in a collection. In medical imaging, positive (labeled) data correspond to images of tissue affected by a disease, while the remaining available images of the same tissue comprise the unlabeled data. Protein identification using mass spectrometry is another variant of such a general problem.
In this work, we propose application-specific machine learning algorithms to address this problem. The reliable identification of proteins from mixtures using mass spectrometry would provide an important tool in both biomedical research and clinical practice.
Title: Matrix Approximation for Large-scale Learning
Candidate: Talwalkar, Ameet
Advisor(s): Mohri, Mehryar
Abstract:
Modern learning problems in computer vision, natural language processing, computational biology, and other areas are often based on large data sets of thousands to millions of training instances. However, several standard learning algorithms, such as kernel-based algorithms, e.g., Support Vector Machines, Kernel Ridge Regression, Kernel PCA, do not easily scale to such orders of magnitude. This thesis focuses on sampling-based matrix approximation techniques that help scale kernel-based algorithms to large-scale datasets. We address several fundamental theoretical and empirical questions including:
What approximation should be used? We discuss two common sampling-based methods, providing novel theoretical insights regarding their suitability for various applications and experimental results motivated by this theory. Our results show that one of these methods, the Nystrom method, is superior in the context of large-scale learning.
Do these approximations work in practice? We show the effectiveness of approximation techniques on a variety of problems. In the largest study to-date for manifold learning, we use the Nystrom method to extract low-dimensional structure from high-dimensional data to effectively cluster face images. We also report good empirical results for kernel ridge regression and kernel logistic regression.
How should we sample columns? A key aspect of sampling-based algorithms is the distribution according to which columns are sampled. We study both fixed and adaptive sampling schemes as well as a promising ensemble technique that can be easily parallelized and generates superior approximations, both in theory and in practice.
How well do these approximations work in theory? We provide theoretical analyses of the Nystrom method to understand when this technique should be used. We present guarantees on approximation accuracy based on various matrix properties and analyze the effect of matrix approximation on actual kernel-based algorithms.
This work has important consequences for the machine learning community since it extends to large-scale applications the benefits of kernel-based algorithms. The crucial aspect of this research, involving low-rank matrix approximation, is of independent interest within the field of numerical linear algebra.
Title: Learning Image Decompositions with Hierarchical Sparse Coding
Author(s): Zeiler, Matthew D.; Fergus, Rob
Abstract:
We present a hierarchical model that learns image decompositions via alternating layers of convolutional sparse coding and max pooling. When trained on natural images, the layers of our model capture image information in a variety of forms: low-level edges, mid-level edge junctions, high-level object parts and complete objects. To build our model we rely on a novel inference scheme that ensures each layer reconstructs the input, rather than just the output of the layer directly beneath, as is common with existing hierarchical approaches. This scheme makes it possible to robustly learn multiple layers of representation and we show a model with 4 layers, trained on images from the Caltech-101 dataset. We use our model to produce image decompositions that, when used as input to standard classification schemes, give a significant performance gain over low-level edge features and yield an overall performance competitive with leading approaches.
Title: Factor Graphs for Relational Regression
Candidate: Chopra, Sumit
Advisor(s): LeCun, Yann
Abstract:
Inherent in many interesting regression problems is a rich underlying inter-sample "Relational Structure". In these problems, the samples may be related to each other in ways such that the unknown variables associated with any sample not only depends on its individual attributes, but also depends on the variables associated with related samples. One such problem, whose importance is further emphasized by the present economic crises, is understanding real estate prices. The price of a house clearly depends on its individual attributes, such as, the number of bedrooms. However, the price also depends on the neighborhood in which the house lies and on the time period in which it was sold. This effect of neighborhood and time on the price is not directly measurable. It is merely reflected in the prices of other houses in the vicinity that were sold around the same time period. Uncovering these spatio-temporal dependencies can certainly help better understand house prices, while at the same time improving prediction accuracy.
Problems of this nature fall in the domain of "Statistical Relational Learning". However the drawback of most models proposed so far is that they cater only to classification problems. To this end, we propose "relational factor graph" models for doing regression in relational data. A single factor graph is used to capture, one, dependencies among individual variables of sample, and two, dependencies among variables associated with multiple samples. The proposed models are capable of capturing hidden inter-sample dependencies via latent variables, and also permits non-linear log-likelihood functions in parameter space, thereby allowing considerably more complex architectures. Efficient inference and learning algorithms for relational factor graphs are proposed. The models are applied to predicting the prices of real estate properties and for constructing house price indices. The relational aspect of the model accounts for the hidden spatio-temporal influences on the price of every house. Experiments show that one can achieve considerably superior performance by identifying and using the underlying spatio-temporal structure associated with the problem. To the best of our knowledge this is the first work in the direction of relational regression and is also the first work in constructing house price indices by simultaneously accounting for the spatio-temporal effects on house prices using large-scale industry standard data set.
Title: Hybrid Domain Decomposition Algorithms for Compressible and Almost Incompressible Elasticity
Author(s): Dohrmann, Clark R.; Widlund, Olof B.
Abstract:
Overlapping Schwarz methods are considered for mixed finite element approximations of linear elasticity, with discontinuous pressure spaces, as well as for compressible elasticity approximated by standard conforming finite elements. The coarse components of the preconditioners are based on %spaces, with a fixed number of degrees of freedom per subdomain, spaces, with a number of degrees of freedom per subdomain which is uniformly bounded, and which are similar to those previously developed for scalar elliptic problems and domain decomposition methods of iterative substructuring type, i.e., methods based on non-overlapping decompositions of the domain. The local components of the new preconditioners are based on solvers on a set of overlapping subdomains.
Title: Numerical Estimation of the Second Largest Eigenvalue of a Reversible Markov Transition Matrix
Candidate: Gade, Kranthi
Advisor(s): Goodman, Jonathan
Abstract:
We discuss the problem of finding the second largest eigenvalue of an operator that defines a reversible Markov chain. The second largest eigenvalue governs the rate at which the statistics of the Markov chain converge to equilibrium. Scientific applications include understanding the very slow dynamics of some models of dynamic glass. Applications in computing include estimating the rate of convergence of Markov chain Monte Carlo algorithms.
Most practical Markov chains have state spaces so large that direct or even iterative methods from linear algebra are inapplicable. The size of the state space, which is the dimension of the eigenvalue problem, grows exponentially with the system size. This makes it impossible to store a vector (for sparse methods), let alone a matrix (for dense methods). Instead, we seek a method that uses only time correlation from samples produced from the Markov chain itself.
In this thesis, we propose a novel Krylov subspace type method to estimate the second eigenvalue from the simulation data of the Markov chain using test functions which are known to have good overlap with the slowest mode. This method starts with the naive Rayleigh quotient estimate of the test function and refines it to obtain an improved estimate of the second eigenvalue. We apply the method to a few model problems and the estimate compares very favorably with the known answer. We also apply the estimator to some Markov chains occuring in practice, most notably in the study of glasses. We show experimentally that our estimator is more accurate and stable for these problems compared to the existing methods.
Title: 2D-Centric Interfaces and Algorithms for 3D Modeling
Candidate: Gingold, Yotam
Advisor(s): Zorin, Denis
Abstract:
The creation of 3D models is a fundamental task in computer graphics. The task is required by professional artists working on movies, television, and games, and desired by casual users who wish to make their own models for use in virtual worlds or as a hobby.
In this thesis, we consider approaches to creating and editing 3D models that minimize the user's thinking in 3D. In particular, our approaches do not require the user to manipulate 3D positions in space or mentally invert complex 3D-to-2D mappings. We present interfaces and algorithms for the creation of 3D surfaces, for texturing, and for adding small-to-medium scale geometric detail.
First, we present a novel approach for texture placement and editing based on direct manipulation of textures on the surface. Compared to conventional tools for surface texturing, our system combines UV-coordinate specification and texture editing into one seamless process, reducing the need for careful initial design of parameterization and providing a natural interface for working with textures directly on 3D surfaces.
Second, we present a system for free-form surface modeling that allows a user to modify a shape by changing its rendered, shaded image using stroke-based drawing tools. A new shape, whose rendered image closely approximates user input, is c omputed using an efficient and stable surface optimization procedure. We demonstrate how several types of free-form surface edits which may be difficult to cast in terms of standard deformation approaches can be easily performed using our system.
Third, we present a single-view 2D interface for 3D modeling based on the idea of placing 2D primitives and annotations on an existing, pre-made sketch or image. Our interface frees users to create 2D sketches from arbitrary angles using their preferred tool---including pencil and paper---which they then "describe" using our tool to create a 3D model. Our primitives are manipulated with persistent, dynamic handles, and our annotations take the form of markings commonly used in geometry textbooks.
Title: Proximity problems for point sets residing in spaces with low doubling dimension
Candidate: Gottlieb, Lee-Ad
Advisor(s): Cole, Richard
Abstract:
In this thesis we consider proximity problems on point sets. Proximity problems arise in all fields of computer science, with broad application to computation geometry, machine learning, computational biology, data mining and the like. In particular, we will consider the problems of approximate nearest neighbor search, and dynamic maintenance of a spanner for a point set.
It has been conjectured that all algorithms for these two problems suffer from the "curse of dimensionality," which means that their run time grow exponentially with the dimension of the point set. To avoid this undesirable growth, we consider point sets that occupy a doubling dimension lambda. We first present a dynamic data structure that uses linear space and supports a (1+e)-approximate nearest neighbor search of the point set. We then extend this algorithm to allow the dynamic maintenance of a low degree (1+e)-spanner for the point set. The query and update time of these structures are exponential in lambda (as opposed to exponential in the dimension); when lambda is small, this provides a significant spead-up over known algorithms, and when lambda is constant then these run times are optimal up to a constant. Even when no assumptions are made on lambda, the query and update times of the neighest neighbor search structure match the best known run times for approximate nearest neighbor search (up to a constant multiple in lambda). Further, the stretch of the spanner is optimal, and its update times exceed all previously known algorithms.
Title: Creativity Support for Computational Literature
Candidate: Howe, Daniel
Advisor(s): Perlin, Ken
Abstract:
The creativity support community has a long history of providing valuable tools to artists and designers. Similarly, creative digital media practice has proven a valuable pedagogical strategy for teaching core computational ideas. Neither strain of research has focused on the domain of literary art however, instead targeting visual, and aural media almost exclusively.
To address this situation, this thesis presents a software toolkit created specifically to support creativity in computational literature. Two primary hypotheses direct the bulk of the research presented: first, that it is possible to implement effective creativity support tools for literary art given current resource constraints; and second, that such tools, in addition to facilitating new forms of literary creativity, will provide unique opportunities for computer science education.
Designed both for practicing artists and for pedagogy, the research presented directly addresses impediments to participation in the field for a diverse range of users and provides an end-to-end solution for courses attempting to engage the creative faculties of computer science students, and to introduce a wider demographic--from writers, to digital artists, to media and literary theorists --to procedural literacy and computational thinking.
The tools and strategies presented have been implemented, deployed, and iteratively refined in a real-world contexts over the past three years. In addition to their use in large-scale projects by contemporary artists, they have provided effective support for multiple iterations of 'Programming for Digital Art & Literature', a successful inter-disciplinary computer science course taught by the author.
Taken together, this thesis provides a novel set of tools for a new domain, and demonstrates their real-world efficacy in providing both creativity and pedagogical support for a diverse and emerging population of users.
Title: A numerical method for simulating the dynamics of 3D axisymmetric vesicles suspended in viscous flows
Author(s): K. Veerapaneni, Shravan; Gueyerer, Denis; Biros, George; Zorin, Denis
Abstract:
We extend "A boundary integral method for simulating the dynamics of inextensible vesicles suspended in a viscous fluid in 2D", Veerapaneni et al. Journal of Computational Physics, 228(7), 2009 to the case of three dimensional axisymmetric vesicles of spherical or toroidal topology immersed in viscous flows. Although the main components of the algorithm are similar in spirit to the 2D case.spectral approximation in space, semi-implicit time-stepping scheme.the main differences are that the bending and viscous force require new analysis, the linearization for the semi-implicit schemes must be rederived, a fully implicit scheme must be used for the toroidal topology to eliminate a CFL-type restriction, and a novel numerical scheme for the evaluation of the 3D Stokes single-layer potential on an axisymmetric surface is necessary to speed up the calculations. By introducing these novel components, we obtain a time-scheme that experimentally is unconditionally stable, has low cost per time step, and is third-order accurate in time. We present numerical results to analyze the cost and convergence rates of the scheme. To verify the solver, we compare it to a constrained variational approach to compute equilibrium shapes that does not involve interactions with a viscous fluid. To illustrate the applicability of method, we consider a few vesicle-flow interaction problems: the sedimentation of a vesicle, interactions of one and three vesicles with a background Poiseuille flow.
Title: A Hybrid Domain Decomposition Method and its Applications to Contact Problems
Author(s): Lee, Jungho
Abstract:
Our goal is to solve nonlinear contact problems. We consider bodies in contact with each other divided into subdomains, which in turn are unions of elements. The contact surface between the bodies is unknown a priori, and we have a nonpen-etration condition between the bodies, which is essentially an inequality constraint. We choose to use an active set method to solve such problems, which has both outer iterations in which the active set is updated, and inner iterations in which a (linear) minimization problem is solved on the current active face. In the first part of this dissertation, we review the basics of domain decomposition methods. In the second part, we consider how to solve the inner minimization problems. Using an approach based purely on FETI algorithms with only Lagrange multipliers as unknowns, as has been developed by the engineering community, does not lead to a scalable algorithm with respect to the number of subdomains in each body. We prove that such an algorithm has a condition number estimate which depends linearly on the number of subdomains across a body; numerical experiments suggest that this is the best possible bound. We also consider a new method based on the saddle point formulation of the FETI methods with both displacement vectors and Lagrange multipliers as unknowns. The resulting system is solved with a block-diagonal preconditioner which combines the one-level FETIand the BDDC methods. This approach allows the use of inexact solvers. We show that this new method is scalable with respect to the number of subdomains, and that its convergence rate depends only logarithmically on the number of degrees of freedom of the subdomains and bodies. In the last part of this dissertation, a model contact problem is solved by two approaches. The first one is a nonlinear algorithm which combines an active set method and the new method of Chapter 4. We also present a novel way of finding an initial active set. The second one uses the SMALBE algorithm, developed by Dostal et al. We show that the former approach has advantages over the latter.
Title: Efficient Systems Biology Algorithms for Biological Networks over Multiple Time-Scales: From Evolutionary to Regulatory Time
Candidate: Mitrofanova, Antonina
Advisor(s): Mishra, Bud
Abstract:
Recently, Computational Biology has emerged as one of the most exciting areas of computer science research, not only because of its immediate impact on many biomedical applications, (e.g., personalized medicine, drug and vaccine discovery, tools for diagnostics and therapeutic interventions, etc.), but also because it raises many new and interesting combinatorial and algorithmic questions, in the process. In this thesis, we focus on robust and efficient algorithms to analyze biological networks, primarily targeting protein networks, possibly the most fascinating networks in computational biology in terms of their structure, evolution and complexity, as well as because of their role in various genetic and metabolic diseases.
Classically, protein networks have been studied statically, i.e., without taking into account time-dependent metamorphic changes in network topology and functionality. In this work, we introduce new analysis techniques that view protein networks as being dynamic in nature, evolving over time, and diverse in regulatory patterns at various stages of the system development. Our analysis is capable of dealing with multiple time-scales: ranging from the slowest time-scale corresponding to evolutionary time between species, speeding up to inter-species pathway evolution time, and finally, moving to the other extreme at the cellular developmental time-scale.
We also provide a new method to overcome limitations imposed by corrupting effects of experimental noise (e.g., high false positive and false negative rates) in Yeast Two-Hybrid (Y2H) networks, which often provide primary data for protein complexes. Our new combinatorial algorithm measures connectivity between proteins in Y2H network not by edges but by edge-disjoint paths, which reflects pathway evolution better within single specie network. This algorithm has been shown to be robust against increasing false positives and false negatives, as estimated using variation of information and separation measures.
In addition, we have devised a new way to incorporate evolutionary information in order to significantly improve classification of proteins, especially those isolated in their own networks or surrounded by poorly characterized neighbors. In our method, the networks of two (or more) species are joined by edges of high sequence similarity so that protein-homologs of different species can exchange information and acquire new and improved functional associations.
Finally, we have integrated many of these techniques into one tool to create a novel analysis of malaria parasite P. falciparum's life-cycle at the scale of reaction-time, single cell level, and encompassing its entire inter-erythrocytic developmental cycle (IDC). Our approach allows connecting time-course gene expression profiles of consecutive IDC stages in order to assign functions to un-annotated Malaria proteins and predict potential targets for vaccine and drug development.
Title: Detecting, modeling and rendering complex configurations of curvilinear features
Candidate: Parilov, Evgueni
Advisor(s): Zorin, Denis
Abstract:
Curvilinear features allow one to represent a variety of real world regular patterns like honeycomb tiling as well as very complicated random patterns like networks of furrows on the surface of the human skin. We have developed a set of methods and new data representations for solving key problems related to curvilinear features, which include robust detection of intricate networks of curvilinear features from digital images, GPU-based sharp rendering of fields with curvilinear features, and a parametric synthesis approach to generate systems of curvilinear features with desirable local configurations and global control.
The existing edge-detection techniques may underperform in the presence of noise, usually do not link the detected edge points into chains, often fail on complex structures, heavily depend on initial guess, and assume significant manual phase. We have developed a technique based on active contours, or snakes, which avoids manual initial positioning of the snakes and can detect large networks of curves with complex junctions without user guidance.
The standard bilinear interpolation of piecewise continuous fields results in unwanted smoothing along the curvilinear discontinuities. Spatially varying features can be best represented as a function of the distance to the discontinuity curves and its gradient. We have developed a real-time, GPU-based method for unsigned distance function field and its gradient field interpolation which preserves discontinuity feature curves, represented by quadratic Bezier curves, with minimal restriction on their topology.
Detail features are very important visual clues which make computer-generated imagery look less artificial. Instead of using sample-based synthesis technique which lacks user control on features usually producing gaps in features or breaking feature coherency, we have explored an alternative approach of generating features using random fibre processes. We have developed a Gibbs-type random process of linear fibres based on local fibre interactions. It allows generating non-stationary curvilinear networks with some degree of regularity, and provides an intuitive set of parameters which directly defines fibre local configurations and global pattern of fibres.
For random systems of linear fibres which approximately form two orthogonal dominant orientation fields, we have adapted a streamline placement algorithm which converts such systems into overlapping random sets of coherent smooth curves.
Title: Unsupervised Learning of Feature Hierarchies
Candidate: Ranzato, Marc'Aurelio
Advisor(s): LeCun, Yann
Abstract:
The applicability of machine learning methods is often limited by the amount of available labeled data, and by the ability (or inability) of the designer to produce good internal representations and good similarity measures for the input data vectors.
The aim of this thesis is to alleviate these two limitations by proposing algorithms to learn good internal representations, and invariant feature hierarchies from unlabeled data. These methods go beyond traditional supervised learning algorithms, and rely on unsupervised, and semi-supervised learning.
In particular, this work focuses on ''deep learning'' methods, a set of techniques and principles to train hierarchical models. Hierarchical models produce feature hierarchies that can capture complex non-linear dependencies among the observed data variables in a concise and efficient manner. After training, these models can be employed in real-time systems because they compute the representation by a very fast forward propagation of the input through a sequence of non-linear transformations.
When the paucity of labeled data does not allow the use of traditional supervised algorithms, each layer of the hierarchy can be trained in sequence starting at the bottom by using unsupervised or semi-supervised algorithms. Once each layer has been trained, the whole system can be fine-tuned in an end-to-end fashion. We propose several unsupervised algorithms that can be used as building block to train such feature hierarchies. We investigate algorithms that produce sparse overcomplete representations and features that are invariant to known and learned transformations. These algorithms are designed using the Energy-Based Model framework and gradient-based optimization techniques that scale well on large datasets. The principle underlying these algorithms is to learn representations that are at the same time sparse, able to reconstruct the observation, and directly predictable by some learned mapping that can be used for fast inference in test time.
With the general principles at the foundation of these algorithms, we validate these models on a variety of tasks, from visual object recognition to text document classification and retrieval.
Title: Learning least squares estimators without assumed priors or supervision
Author(s): Raphan, Martin; Simoncelli, Eero P.
Abstract:
The two standard methods of obtaining a least-squares optimal estimator are (1) Bayesian estimation, in which one assumes a prior distribution on the true values and combines this with a model of the measurement process to obtain an optimal estimator, and (2) supervised regression, in which one optimizes a parametric estimator over a training set containing pairs of corrupted measurements and their associated true values. But many real-world systems do not have access to either supervised training examples or a prior model. Here, we study the problem of obtaining an optimal estimator given a measurement process with known statistics, and a set of corrupted measurements of random values drawn from an unknown prior. We develop a general form of nonparametric empirical Bayesian estimator that is written as a direct function of the measurement density, with no explicit reference to the prior. We study the observation conditions under which such "prior-free" estimators may be obtained, and we derive specific forms for a variety of different corruption processes. Each of these prior-free estimators may also be used to express the mean squared estimation error as an expectation over the measurement density, thus generalizing Stein's unbiased risk estimator (SURE) which provides such an expression for the additive Gaussian noise case. Minimizing this expression over measurement samples provides an "unsupervised regression" method of learning an optimal estimator from noisy measurements in the absence of clean training data. We show that combining a prior-free estimator with its corresponding unsupervised regression form produces a generalization of the "score matching" procedure for parametric density estimation, and we develop an incremental form of learning for estimators that are written as a linear combination of nonlinear kernel functions. Finally, we show through numerical simulations that the convergence of these estimators can be comparable to their supervised or Bayesian counterparts.
Title: Plinkr: an Application of Semantic Search
Candidate: Scott, John
Advisor(s): Shasha, Dennis
Abstract:
Plinkr extends and enriches traditional keyword search with semantic search technology. Specifically, Plinkr facilitates the process of discovering the intersection of information between two subjects. This intersection represents what the subjects have in common and thus effectively captures the relationships between them. This is accomplished by semantically tagging and scoring entities that are contained within various keyword searches. The most relevant entities are thus abstracted and presented as metadata which can be explored to highlight the most pertinent content.
Title: Search Problems for Speech and Audio Sequences
Candidate: Weinstein, Eugene
Advisor(s): Mohri, Mehryar
Abstract:
The modern proliferation of very large audio, video, and biological databases has created a need for the design of effective methods for indexing and searching highly variable or uncertain data. Classical search and indexing algorithms deal with clean or perfect input sequences. However, an index created from speech transcriptions is marked with errors and uncertainties stemming from the use of imperfect statistical models in the speech recognition process. Similarly, automatic transcription of music, such as assigning a sequence of notes to represent a stream of music audio, is prone to errors. How can we generalize search and indexing algorithms to deal with such uncertain inputs?
This thesis presents several novel algorithms, analyses, and general techniques and tools for effective indexing and search that not only tolerate but actually exploit this uncertainty. In particular, it develops an algorithmic foundation for music identification, or content-based music search; presents novel automata-theoretic results applicable generally to a variety of search and indexing tasks; and describes new algorithms for topic segmentation, or automatic splitting of speech streams into topic-coherent segments.
We devise a new technique for music identification in which each song is represented by a distinct sequence of music sounds, called "music phonemes." In our approach, we learn the set of music phonemes, as well as a unique sequence of music phonemes characterizing each song, from training data using an unsupervised algorithm. We also propose a novel application of factor automata to create a compact mapping of music phoneme sequences to songs. Using these techniques, we construct an efficient and robust music identification system for a large database of songs.
We further design new algorithms for compact indexing of uncertain inputs based on suffix and factor automata and give novel theoretical guarantees for their space requirements. Suffix automata and factor automata represent the set of all suffixes or substrings of a set of strings, and are used in numerous indexing and search tasks, including the music identification system just mentioned. We show that the suffix automaton or factor automaton of a set of strings U has at most 2Q-2 states, where Q is the number of nodes of a prefix-tree representing the strings in U, a significant improvement over previous work. We also describe a matching new linear-time algorithm for constructing the suffix automaton S or factor automaton F of U in time O(|S|).
We also define a new quality measure for topic segmentation systems and design a discriminative topic segmentation algorithm for speech inputs, thus facilitating effective indexation of spoken audio collections. The new quality measure improves on previously used criteria and is correlated with human judgment of topic-coherence. Our segmentation algorithm uses a novel general topical similarity score based on word co-occurrence statistics. This new algorithm outperforms previous methods in experiments over speech and text streams. We further demonstrate that the performance of segmentation algorithms can be improved by using a lattice of competing hypotheses over the speech stream rather than just the one-best hypothesis as input.
Title: Body Signature Recognition
Author(s): Williams, George; Bregler, Christoph; Hackney, Peggy; Rosenthal, Sally; McDowall, Ian; Smolskiy, Kirill
Abstract:
This paper describes a new visual representation of motion that is used to learn and classify body language - what we call .body signatures. - of people while they are talking. We applied this technique to several hours of internet videos and television broadcasts that include US politicians and leaders from Germany, France, Iran, Russia, Pakistan, and India, and public figures such as the Pope, as well as numerous talk show hosts and comedians. Dependent on the complexity of the task, we show up to 80% recognition performance and clustering into broader body language categories.
Title: Using Application-Domain Knowledge in the Runtime Support of Multi-Experiment Computational Studies
Candidate: Yau, Siu-Man
Advisor(s): Karamcheti, Vijay; Zorin, Denis
Abstract:
Multi-Experiment Studies (MESs) is a type of computational study in which the same simulation software is executed multiple times, and the result of all executions need to be aggregated to obtain useful insight. As computational simulation experiments become increasingly accepted as part of the scientific process, the use of MESs is becoming more wide-spread among scientists and engineers.
MESs present several challenging requirements on the computing system. First, many MESs need constant user monitoring and feedback, requiring simultaneous steering of multiple executions of the simulation code. Second, MESs can comprise of many executions of long-running simulations; the sheer volume of computation can make them prohibitively long to run.
Parallel architecture offer an attractive computing platform for MESs. Low-cost, small-scale desktops employing multi-core chips allow wide-spread dedicated local access to parallel computation power, offering more research groups an opportunity to achieve interactive MESs. Massively-parallel, high-performance computing clusters can afford a level of parallelism never seen before, and present an opportunity to address the problem of computationally intensive MESs.
However, in order to fully leverage the benefits of parallel architectures, the traditional parallel systems' view has to be augmented. Existing parallel computing systems often treat each execution of the software as a black box, and are prevented from viewing an entire computational study as a single entity that must be optimized for.
This dissertation investigates how a parallel system can view MESs as an end-to-end system and leverage the application-specific properties of MESs to address its requirements. In particular, the system can 1) adapt its scheduling decisions to the overall goal of an MES to reduce the needed computation, 2) simultaneously aggregate results from, and disseminate user actions to, multiple executions of the software to enable simultaneous steering, 3) store reusable information across executions of the simulation software to reduce individual run-time, and 4) adapt its resource allocation policies to the MES's properties to improve resource utilization.
Using a test bed system called SimX and four example MESs across different disciplines, this dissertation shows that the application-aware MES-level approach can achieve multi-fold to multiple orders-of-magnitude improvements over the traditional simulation-level approach.
Title: Ensuring Correctness of Compiled Code
Candidate: Zaks, Ganna
Advisor(s): Pnueli, Amir
Abstract:
Traditionally, the verification effort is applied to the abstract algorithmic descriptions of the underlining software. However, even well understood protocols such as Peterson's protocol for mutual exclusion, whose algorithmic description takes only half a page, have published implementations that are erroneous. Furthermore, the semantics of the implementations can be altered by optimizing compilers, which are very large applications and, consequently, are bound to have bugs. Thus, it is highly desirable to ensure the correctness of the compiled code especially in safety critical and high-assurance software. This dissertation describes two alternative approaches that bring us closer to solving the problem.
First, we present CoVaC - a deductive framework for proving program equivalence and its application to automatic verification of transformations performed by optimizing compilers. To leverage the existing program analysis techniques, we reduce the equivalence checking problem to analysis of one system - a cross-product of the two input programs. We show how the approach can be effectively used for checking equivalence of single-threaded programs that are structurally similar. Unlike the existing frameworks, our approach accommodates absence of compiler annotations and handles most of the classical intraprocedural optimizations such as constant folding, reassociation, common subexpression elimination, code motion, dead code elimination, branch optimizations, and others. In addition, we have developed rules for translation validation of interprocedural optimizations, which can be applied when compiler annotations are available.
The second contribution is the pancam framework for verifying multi-threaded C programs. Pancam first compiles a multithreaded C program into optimized bytecode format. The framework relies on Spin, an existing explicit state model checker, to orchestrate the program's state space search. However, the program transitions and states are computed by the pancam bytecode interpreter. A feature of our approach is that not only pancam checks the actual implementation, but it can also check the code after compiler optimizations. Pancam addresses the state space explosion problem by allowing users to define data abstraction functions and to constrain the number of allowed context switches. We also describe a partial order reduction method that reduces context switches using dynamic knowledge computed on-the-fly, while being sound for both safety and liveness properties.
Title: General Algorithms for Testing the Ambiguity of Finite Automata
Author(s): Allauzen, Cyril; Mohri, Mehryar; Rastogi, Ashish
Abstract:
This paper presents efficient algorithms for testing the finite, polynomial, and exponential ambiguity of finite automata with $\epsilon$-transitions. It gives an algorithm for testing the exponential ambiguity of an automaton $A$ in time $O(|A|_E2)$, and finite or polynomial ambiguity in time $O(|A|_E3)$. These complexities significantly improve over the previous best complexities given for the same problem. Furthermore, the algorithms presented are simple and are based on a general algorithm for the composition or intersection of automata. We also give an algorithm to determine the degree of polynomial ambiguity of a finite automaton $A$ that is polynomially ambiguous in time $O(|A|_E3)$. Finally, we present an application of our algorithms to an approximate computation of the entropy of a probabilistic automaton.
Title: Competitive Hybridization Model
Author(s): Cherepinsky, Vera; Hashmi, Ghazala; Seul, Michael; Mishra, Bud
Abstract:
Microarray technology, in its simplest form, allows one to gather abundance data for target DNA molecules, associated with genomes or gene-expressions, and relies on hybridizing the target to many short probe oligonucleotides arrayed on a surface. While for such multiplexed reactions conditions are optimized to make the most of each individual probe-target interaction, subsequent analysis of these experiments is based on the implicit assumption that a given experiment gives the same result regardless of whether it was conducted in isolation or in parallel with many others. It has been discussed in the literature that this assumption is frequently false, and its validity depends on the types of probes and their interactions with each other. We present a detailed physical model of hybridization as a means of understanding probe interactions in a multiplexed reaction. The model is formulated as a system of ordinary di.erential equations (ODE.s) describing kinetic mass action and conservation-of-mass equations completing the system.
We examine pair-wise probe interactions in detail and present a model of .competition. between the probes for the target.especially, when target is in short supply. These e.ects are shown to be predictable from the a.nity constants for each of the four probe sequences involved, namely, the match and mismatch for both probes. These a.nity constants are calculated from the thermodynamic parameters such as the free energy of hybridization, which are in turn computed according to the nearest neighbor (NN) model for each probe and target sequence.
Simulations based on the competitive hybridization model explain the observed variability in the signal of a given probe when measured in parallel with di.erent groupings of other probes or individually. The results of the simulations are used for experiment design and pooling strategies, based on which probes have been shown to have a strong e.ect on each other.s signal in the in silico experiment. These results are aimed at better design of multiplexed reactions on arrays used in genotyping (e.g., HLA typing, SNP or CNV detection, etc.) and mutation analysis (e.g., cystic .brosis, cancer, autism, etc.).
Title: Friendshare: A decentralized, consistent storage repository for collaborative file sharing
Candidate: Chiang, Frank
Advisor(s): Li, Jinyang
Abstract:
Data sharing has become more and more collaborative with the rise of Web 2.0, where multiple writers jointly write and organize the content in a repository. Current solutions use a centralized entity, such as Wikipedia or Google Groups, to serve the data. However, centralized solutions may be undesirable due to privacy concerns and censorship, which are problems that can be alleviated by switching to decentralized solutions.
The challenge of building a decentralized collaborative repository is achieving high data availability, durability, and consistency. Attaining these goals is difficult because peer nodes have limited bandwidth and storage space, low availability, and the repository has high membership churn.
This thesis presents Friendshare, a decentralized multiple-writer data repository. Separating the metadata from the data allows for efficient metadata replication across privileged admin nodes, thus increasing availability and durability. The primary commit scheme, where a primary node is responsible for determining the total order of writes in the repository, is employed to ensure eventual consistency. If the primary leaves the system unexpectedly, the remaining admin nodes run Paxos, a consensus protocol, to elect a new primary.
The Paxos protocol requires high node availability in order to be run efficiently, a criteria that is rarely met in typical peer-to-peer networks. To rectify this problem, we offer two optimizations to improve Paxos performance in low availability environments.
Friendshare has been implemented and deployed to gather real-world statistics. To offer theoretical predictions, we built a simulator to demonstrate the performance and service availability of Friendshare at various node online percentages. In addition, we show the performance improvements of our Paxos optimizations in comparison with the basic Paxos protocol.
Title: Factor Graphs for Relational Regression
Author(s): Chopra, Sumit; Thampy, Trivikaraman; Leahy, John; Caplin, Andrew; LeCun, Yann
Abstract:
Traditional methods for supervised learning involve treating the input data as a set of independent, identically distributed samples. However, in many situations, the samples are related in such a way that variables associated with one sample depend on other samples. We present a new form of relational graphical model that, in addition to capturing the dependence of the output on sample specific features, can also capture hidden relationships among samples through a non-parametric latent manifold. Learning in the proposed graphical model involves simultaneously learning the non-parametric latent manifold along with a non-relational parametric model. Efficient inference algorithms are introduced to accomplish this task. The method is applied to the prediction of house prices. A non-relational model predicts an ``intrinsic" price of the house which depends only on its individual characteristics, and a relational model estimates a hidden surface of ``desirability'' coefficients which links the price of a house to that of similar houses in the neighborhood.
Title: Verification of Transactional Memories and Recursive Programs
Candidate: Cohen, Ariel
Advisor(s): Pnueli, Amir
Abstract:
Transactional memory is a programming abstraction intended to simplify the synchronization of conflicting concurrent memory accesses without the difficulties associated with locks. In the first part of this thesis we provide a framework and tools that allow to formally verify that a transactional memory implementation satisfies its specification. First we show how to specify transactional memory in terms of admissible interchanges of transaction operations, and give proof rules for showing that an implementation satisfies its specification. We illustrate how to verify correctness, first using a model checker for bounded instantiations, and subsequently by using a theorem prover, thus eliminating all bounds. We provide a mechanical proof of the soundness of the verification method, as well as mechanical proofs for several implementations from the literature, including one that supports non-transactional memory accesses.
Procedural programs with unbounded recursion present a challenge to symbolic model-checkers since they ostensibly require the checker to model an unbounded call stack. In the second part of this thesis we present a method for model-checking safety and liveness properties over procedural programs. Our method performs by first augmenting a concrete procedural program with a well founded ranking function, and then abstracting the Procedural programs with unbounded recursion present a challenge to symbolic model-checkers since they ostensibly require the checker to model an unbounded call stack. In the second part of this thesis we present a method for model-checking safety and liveness properties over procedural programs. Our method performs by first augmenting a concrete procedural program with a well founded ranking function, and then abstracting the augmented program by a finitary state abstraction. Using procedure summarization the procedural abstract program is then reduced to a finite-state system, which is model checked for the property.
Title: Pointer Analysis, Conditional Soundness, and Proving the Absence of Errors
Author(s): Conway, Christopher L.; Dams, Dennis; Namjoshi, Kedar S.; Barrett, Clark
Abstract:
It is well known that the use of points-to information can substantially improve the accuracyof a static program analysis. Commonly used algorithms for computing points-to information are known to be sound only for memory-safe programs. Thus, it appears problematic to utilize points-to information to verify the memory safety property without giving up soundness. We show that a sound combination is possible, even if the points-to information is computed separately and only conditionally sound. This result is based on a refined statement of the soundness conditions of points-to analyses and a general mechanism for composing conditionally sound analyses.
Title: STUMP: Stereo Correspondence in the Cyclopean Eye under Belief Propagation
Candidate: Distler, George
Advisor(s): Geiger, Davi
Abstract:
The human visual system sees at any moment a static scene in three dimensions. This 3D view of the world is acquired by two images, one from the left eye and the other by the right eye. Fusing the left and right stereo pair of images yields a single cyclopean view portraying depth. Stereo vision can be applied to the field of computer vision via calibrated stereo cameras to capture the left and right images. Given a stereo pair of images, one can compute the field of depth via a stereo correspondence algorithm. We present a new approach to computing the disparity (depth) by means the STUMP algorithm.
The STUMP algorithm presents a solution to the stereo correspondence problem. We propose to solve the problem of discontinuities in disparity within epipolar lines by modeling geometric constraints of smooth, tilted, and occluded surfaces as well as unicity and opaqueness. Our algorithm runs upon a framework built upon the BP-TwoGraphs belief propagation estimation [17]. As a result, we provide a disparity map in the cyclopean coordinate system determined by a probability distribution computed in polynomial time.
Title: An Overlapping Schwarz Algorithm for Almost Incompressible Elasticity
Author(s): Dohrmann, Clark R.; Widlund, Olof B.
Abstract:
Overlapping Schwarz methods are extended to mixed finite element approximations of linear elasticity which use discontinuous pressure spaces. The coarse component of the preconditioner is based on a low-dimensional space previously developed for scalar elliptic problems and a domain decomposition method of iterative substructuring type, i.e., a method based on non-overlapping decompositions of the domain, while the local components of the preconditioner are based on solvers on a set of overlapping subdomains.
A bound is established for the condition number of the algorithm which grows in proportion to the square of the logarithm of the number of degrees of freedom in individual subdomains and the third power of the relative overlap between the overlapping subdomains, and which is independent of the Poisson ratio as well as jumps in the Lam\'e parameters across the interface between the subdomains. A positive definite reformulation of the discrete problem makes the use of the standard preconditioned conjugate gradient method straightforward. Numerical results, which include a comparison with problems of compressible elasticity, illustrate the findings.
Title: Learning Long-Range Vision for an Offroad Robot
Candidate: Hadsell, Raia
Advisor(s): LeCun, Yann
Abstract:
Teaching a robot to perceive and navigate in an unstructured natural world is a difficult task. Without learning, navigation systems are short-range and extremely limited. With learning, the robot can be taught to classify terrain at longer distances, but these classifiers can be fragile as well, leading to extremely conservative planning. A robust, high-level learning-based perception system for a mobile robot needs to continually learn and adapt as it explores new environments. To do this, a strong feature representation is necessary that can encode meaningful, discriminative patterns as well as invariance to irrelevant transformations. A simple realtime classifier can then be trained on those features to predict the traversability of the current terrain.
One such method for learning a feature representation is discussed in detail in this work. Dimensionality reduction by learning an invariant mapping (DrLIM) is a weakly supervised method for learning a similarity measure over a domain. Given a set of training samples and their pairwise relationships, which can be arbitrarily defined, DrLIM can be used to learn a function that is invariant to complex transformations of the inputs such as shape distortion and rotation.
The main contribution of this work is a self-supervised learning process for long-range vision that is able to accurately classify complex terrain, permitting improved strategic planning. As a mobile robot moves through offroad environments, it learns traversability from a stereo obstacle detector. The learning architecture is composed of a static feature extractor, trained offline for a general yet discriminative feature representation, and an adaptive online classifier. This architecture reduces the effect of concept drift by allowing the online classifier to quickly adapt to very few training samples without overtraining. After experiments with several different learned feature extractors, we conclude that unsupervised or weakly supervised learning methods are necessary for training general feature representations for natural scenes.
The process was developed and tested on the LAGR mobile robot as part of a fully autonomous vision-based navigation system.
Title: Modal Logic, Temporal Models and Neural Circuits: What Connects Them
Author(s): Kleinberg, Samantha; Antoniotti, Marco; Ramakrishnan, Naren; Mishra, Bud
Abstract:
Traditional methods for supervised learning involve treating the input data as a set of independent, identically distributed samples. However, in many situations, the samples are related in such a way that variables associated with one sample depend on other samples. We present a new form of relational graphical model that, in addition to capturing the dependence of the output on sample specific features, can also capture hidden relationships among samples through a non-parametric latent manifold.
Learning in the proposed graphical model involves simultaneously learning the non-parametric latent manifold along with a non-relational parametric model. Efficient inference algorithms are introduced to accomplish this task. The method is applied to the prediction of house prices. A non-relational model predicts an ``intrinsic" price of the house which depends only on its individual characteristics, and a relational model estimates a hidden surface of ``desirability'' coefficients which links the price of a house to that of similar houses in the neighborhood.
Title: Extension of Two-level Schwarz Preconditioners to Symmetric Indefinite Problems
Author(s): Leong, Alan
Abstract:
Two-level overlapping Schwarz preconditioners are extended for use for a class of large, symmetric, indefinite systems of linear algebraic equations. The focus is on an enriched coarse space with additional basis functions built from free space solutions of the underlying partial differential equation. GMRES is used to accelerate the convergence of preconditioned systems. Both additive and hybrid Schwarz methods are considered and reports are given on extensive numerical experiments.
Title: Nonlinear extraction of 'Independent Components' of elliptically symmetric densities using radial Gaussianization
Author(s): Lyu, Siwei; Simoncelli, Eero P.
Abstract:
We consider the problem of efficiently encoding a signal by transforming it to a new representation whose components are statistically independent (also known as factorial). A widely studied family of solutions, generally known as independent components analysis (ICA), exists for the case when the signal is generated as a linear transformation of independent non-Gaussian sources. Here, we examine a complementary case, in which the signal density is non-Gaussian but elliptically symmetric. In this case, no linear transform suffices to properly decompose the signal into independent components, and thus, the ICA methodology fails. We show that a simple nonlinear transformation, which we call radial Gaussianization (RG), provides an exact solution for this case. We then examine this methodology in the context of natural image statistics, demonstrating that joint statistics of spatially proximal coefficients in a multi-scale image representation are better described as elliptical than factorial. We quantify this by showing that reduction in dependency achieved by RG is far greater than that achieved by ICA, for local spatial neighborhoods. We also show that the RG transformation may be closely approximated by divisive normalization transformations that have been used to model the nonlinear response properties of visual neurons, and that have been shown to reduce dependencies between multi-scale image coefficients.
Title: Synthesizing Executable Programs from Requirements
Candidate: Plock, Cory
Advisor(s): Goldberg, Benjamin
Abstract:
Automatic generation of correct software from requirements has long been a ``holy grail'' for system and software development. According to this vision, instead of implementing a system and then working hard to apply testing and verification methods to prove system correctness, a system is rather built correctly by construction. This problem, referred to as synthesis, is undecidable in the general case. However, by restricting the domain to decidable subsets, it is possible to bring this vision one step closer to reality.
The focus of our study is reactive systems, or non-terminating programs that continuously receive input from an external environment and produce output responses. Reactive systems are often safety critical and include applications such as anti-lock braking systems, auto-pilots, and pacemakers. One of the challenges of reactive system design is ensuring that the software meets the requirements under the assumption of unpredictable environment input. The behavior of many of these systems can be expressed as regular languages over infinite strings, a domain in which synthesis has yielded successful results.
We present a method for synthesizing executable reactive systems from formal requirements. The object-oriented requirements language of Live Sequence Charts (LSCs) is considered. We begin by establishing a mapping between various subsets of the language and finite-state formal models. We also consider LSCs which can express time constraints over a dense-time domain. From one of these models, we show how to formulate a winning strategy that is guaranteed to satisfy the requirements, provided one exists. The strategy is realized in the form of a controller which guides the system in choosing only non-violating behaviors. We describe an implementation of this work as an extension of an existing tool called the Play-Engine.
Title: Theory and Algorithms for Modern Machine Learning Problems and an Analysis of Markets
Candidate: Rastogi, Ashish
Advisor(s): Cole, Richard; Mohri, Mehryar
Abstract:
The unprecedented growth of the Internet over the past decade and of data collection, more generally, has given rise to vast quantities of digital information, ranging from web documents and images, genomic databases to a vast array of business customer information. Consequently, it is of growing importance to develop tools and models that enable us to better understand this data and to design data-driven algorithms that leverage this information. This thesis provides several fundamental theoretical and algorithmic results for tackling such problems with applications to speech recognition, image processing, natural language processing, computational biology and web-based algorithms.
Probabilistic automata provide an efficient and compact way to model sequence- oriented data such as speech or web documents. Measuring the similarity of such automata provides a way of comparing the objects they model, and is an essential first step in organizing this type of data. We present algorithmic and hardness results for computing various discrepancies (or dissimilarities) between probabilistic automata, including the relative entropy and the Lp distance; we also give an efficient algorithm to determine if two probabilistic automata are equivalent. In addition, we study the complexity of computing the norms of probabilistic automata.
Organizing and querying large amounts of digitized data such as images and videos is a challenging task because little or no label information is available. This motivates transduction, a setting in which the learning algorithm can leverage unlabeled data during training to improve performance. We present novel error bounds for a family of transductive regression algorithms and validate their usefulness through experiments.
Widespread success of search engines and information retrieval systems has led to large scale collection of rating information which is being used to provide personalized rankings. We examine an alternate formulation of the ranking problem for search engines motivated bythe requirement that in addition to accurately predicting pairwise ordering, ranking systems must also preserve the magnitude of the preferences or the difference between ratings. We present algorithms with sound theoretical properties, and verify their efficacy through experiments.
Finally, price discovery in a market setting can be viewed as an (ongoing) learning problem. Specifically, the problem is to find and maintain a set of prices that balance supply and demand, a core topic in economics. This appears to involve complex implicit and possibly large-scale information transfers. We show that finding equilibrium prices, even approximately, in discrete markets is NP-hard and complement the hardness result with a matching polynomial time approximation algorithm.We also give a new way of measuring the quality of an approximation to equilibrium prices that is based on a natural aggregation of the dissatisfaction of individual market participants.
Title: Measuring biomolecules: an image processing and length estimation pipeline using atomic force microscopy to measure DNA and RNA with high precision
Candidate: Sundstrom, Andrew
Advisor(s): Mishra, Bud
Abstract:
Background. An important problem in molecular biology is to determine the complete transcription profile of a single cell, a snapshot that shows which genes are being expressed and to what degree. Seen in series as a movie, these snapshots would give direct, specific observation of the cell.s regulation behavior. Taking a snapshot amounts to correctly classifying the cell.s ~300 000 mRNA molecules into ~30 000 species, and keeping accurate count of each species. The cell.s transcription profile may be affected by low abundances (1-5 copies) of certain mRNAs; thus, a sufficiently sensitive technique must be employed. A natural choice is to use atomic force microscopy (AFM) to perform single-molecule analysis. Reed et al. ("Single molecule transcription profiling with AFM", Nanotechnology , 18:4 , 2007) developed such an analysis that classifies each mRNA by first multiply cleaving its corresponding synthesized cDNA with a restriction enzyme, then constructing its classification label from ratios of the lengths of its resulting fragments. Thus, they showed the transcription profiling problem reduces to making high-precision measurements of cDNA backbone lengths . correct to within 20-25 bp (6-7.5 nm).
Contribution. We developed an image processing and length estimation pipeline using AFM that can achieve these measurement tolerances. In particular, we developed a biased length estimator using James-Stein shrinkage on trained coefficients of a simple linear regression model, a formulation that subsumes the models we studied.
Methods. First, AFM images were processed to extract molecular objects, skeletonize them, select proper backbone objects from the skeletons, then compute initial lengths of the backbones. Second, a linear regression model was trained on a subset of molecules of known length, namely their computed image feature quantities. Third, the model.s coefficients underwent James-Stein shrinkage to create a biased estimator. Fourth, the trained and tuned model was applied to the image feature quantities computed for each test molecule, giving its final, corrected backbone length.
Results. Training data: one monodisperse set of cDNA molecules of theoretical length 75 nm. Test data: two monodisperse sets of cDNA molecules of unknown length. Corrected distributions of molecular backbone lengths were within 6-7.5 nm from the theoretical lengths of the unknowns, once revealed.
Conclusions. The results suggest our pipeline can be employed in the framework specified by Reed et al. to render single-molecule transcription profiles. The results reveal a high degree of systematic error in AFM measurements that suggests image processing alone is insufficient to achieve a much higher measurement accuracy.
Title: Geometric Modeling with High Order Derivatives
Candidate: Tosun, Elif
Advisor(s): Zorin, Denis
Abstract:
Modeling of high quality surfaces is the core of geometric modeling. Such models are used in many computer-aided design and computer graphics applications. Irregular behavior of higher-order differential parameters of the surface (e.g. curvature variation) may lead to aesthetic or physical imperfections. In this work, we consider approaches to constructing surfaces with high degree of smoothness.
One direction is based on a manifold-based surface definition which ensures well-defined high-order derivatives that can be explicitly computed at any point. We extend previously proposed manifold-based construction to surfaces with piecewise-smooth boundary and explore trade-offs in some elements of the construction. We show that growth of derivative magnitudes with order is a general property of constructions with locally supported basis functions and derive a lower bound for derivative growth and numerically study flexibility of resulting surfaces at arbitrary points.
An alternative direction to using high-order surfaces is to define an approximation to high-order quantities for meshes, with high-order surface implicit. These approximations do not necessarily converge point-wise, but can nevertheless be successfully used to solve surface optimization problems. Even though fourth-order problems are commonly solved to obtain high quality surfaces, in many cases, these formulations may lead to reflection-line and curvature discontinuities. We consider two approaches to further increasing control over surface properties.
The first approach is to consider data-dependent functionals leading to fourth-order problems but with explicit control over desired surface properties. Our fourth-order functionals are based on reflection line behavior. Reflection lines are commonly used for surface interrogation and high-quality reflection line patterns are well-correlated with high-quality surface appearance. We demonstrate how these can be discretized and optimized accurately and efficiently on general meshes.
A more direct approach is to consider a poly-harmonic function on a mesh, such as the fourth-order biharmonic or the sixth-order triharmonic. The biharmonic and the triharmonic equations can be thought of as a linearization of curvature and curvature variation Euler-Lagrange equations respectively. We present a novel discretization for both problems based on the mixed finite element framework and a regularization technique for solving the resulting, highly ill-conditioned systems of equations. We show that this method, compared to more ad-hoc discretizations, has higher degree of mesh independence and yields surfaces of better quality.
Title: An Efficient Reduction of Ranking to Classification
Author(s): Ailon, Nir; Mohri, Mehryar
Abstract:
This paper describes an efficient reduction of the learning problem of ranking to binary classification. As with a recent result of Balcan et al. (2007), the reduction guarantees an average pairwise misranking regret of at most $2r$ using a binary classifier with regret $r$. However, our reduction applies to a broader class of ranking loss functions, admits a simpler proof, and the expected running time complexity of our algorithm in terms of number of calls to a classifier or preference function is improved from $\Omega(n2)$ to $O(n \log n)$. Furthermore, when the top $k$ ranked elements only are required ($k \ll n$), as in many applications in information extraction or search engines, the time complexity of our algorithm can be further reduced to $O(k \log k + n)$. Our reduction and algorithm are thus practical for realistic applications where the number of points to rank exceeds several thousands. Much of our results also extend beyond the bipartite case previously studied.
Title: N-Way Composition of Weighted Finite-State Transducers
Author(s): Allauzen, Cyril; Mohri, Mehryar
Abstract:
Composition of weighted transducers is a fundamental algorithm used in many applications, including for computing complex edit-distances between automata, or string kernels in machine learning, or to combine different components of a speech recognition, speech synthesis, or information extraction system. We present a generalization of the composition of weighted transducers, \emph{$n$-way composition}, which is dramatically faster in practice than the standard composition algorithm when combining more than two transducers. The expected worst-case complexity of our algorithm for composing three transducers $T_1$, $T_2$, and $T_3$\ignore{ depending on the strategy used, is $O(|T_1|_E|T_2|_Q|T_3|_E + |T|)$ or $(|T_1|_Q|T_2|_E|T_3|_Q + |T|)$, } is $O(\min(|T_1|_E|T_2|_Q|T_3|_E, |T_1|_Q|T_2|_E|T_3|_Q) + |T|)$, where $T$ is the result of that composition and $|T_i| = |T_i|_Q + |T_i|_E$ with $|T_i|_Q$ the number of states and $|T_i|_E$ the number of transitions of $T_i$, $i = 1, 2, 3$. In many cases, this significantly improves on the complexity of standard composition. Our algorithm also leads to a dramatically faster composition in practice. Furthermore, standard composition can be obtained as a special case of our algorithm. We report the results of several experiments demonstrating this improvement. These theoretical and empirical improvements significantly enhance performance in the applications already mentioned.
Title: Scaling Data Servers via Cooperative Caching
Candidate: Annapureddy, Siddhartha
Advisor(s): Mazieres, David
Abstract:
In this thesis, we present design techniques -- and systems that illustrate and validate these techniques -- for building data-intensive applications over the Internet. We enable the use of a traditional bandwidth-limited server in these applications. A large number of cooperating users contribute resources such as disk space and network bandwidth, and form the backbone of such applications. The applications we consider fall in one of two categories. The first type provide user-perceived utility in proportion to the data download rates of the participants; bulk data distribution systems is a typical example. The second type are usable only when the participants have data download rates above a certain threshold; video streaming is a prime example.
We built Shark, a distributed file system, to address the first type of applications. It is designed for large-scale, wide-area deployment, while also providing a drop-in replacement for local-area file systems. Shark introduces a novel locality-aware cooperative-caching mechanism, in which clients exploit each other's file caches to reduce load on an origin file server. Shark also enables sharing of data even when it originates from different servers. In addition, Shark clients are mutually distrustful in order to operate in the wide-area. Performance results show that Shark greatly reduces server load and reduces client-perceived latency for read-heavy workloads both in the wide and local areas.
We built RedCarpet, a near-Video-on-Demand (nVoD) system, to address the second type of applications. nVoD allows a user to watch a video starting at any point after waiting for a small setup time. RedCarpet uses a mesh-based peero-peer (P2P) system to provide the nVoD service. In this context, we study the problem of scheduling the dissemination of chunks that constitute a video. We show that providing nVoD is feasible with a combination of techniques that include network coding, avoiding resource starvation for different chunks, and overay topology management algorithms. Our evaluation, using a simulator as well as a prototype, shows that systems that do not optimize in all these dimensions could deliver significantly worse nVoD performance.
Title: Shape Analysis by Abstraction, Augmentation, and Transformation
Candidate: Balaban, Ittai
Advisor(s): Pnueli, Amir; Zuck, Lenore
Abstract:
The goal of shape analysis is to analyze properties of programs that perform destructive updates of linked structures (heaps). This thesis presents an approach for shape analysis based on program augmentation (instrumentation), predicate abstraction, and model checking, that allows for verification of safety and liveness properties (which, for sequential programs, usually corresponds to program invariance and termination).
One of the difficulties in abstracting heap-manipulating programs is devising a decision procedure for a sufficiently expressive logic of graph properties. Since graph reachability (expressible by transitive closure) is not a first order property, the challenge is in showing that a decision procedure exists for a rich enough subset of first order logic with transitive closure.
Predicate abstraction is in general too weak to verify liveness properties. Thus an additional issue dealt with is how to perform abstraction while retaining enough information. The method presented here is domain-neutral, and applies to concurrent programs as well as sequential ones.
Title: Magnitude-Preserving Ranking Algorithms
Author(s): Cortes, Corinna; Mohri, Mehryar; Rastogi, Ashish
Abstract:
This paper studies the learning problem of ranking when one wishes not just to accurately predict pairwise ordering but also preserve the magnitude of the preferences or the difference between ratings, a problem motivated by its crucial importance in the design of search engines, movie recommendation, and other similar ranking systems. We describe and analyze several algorithms for this problem and give stability bounds for their generalization error, extending previously known stability results to non- bipartite ranking and magnitude of preference-preserving algorithms. We also report the results of experiments comparing these algorithms on several datasets and contrast these results with those obtained using an AUC-maximization algorithm.
Title: On the Computation of the Relative Entropy of Probabilistic Automata
Author(s): Cortes, Corinna; Mohri, Mehryar; Rastogi, Ashish; Riley, Michael
Abstract:
We present an exhaustive analysis of the problem of computing the relative entropy of two probabilistic automata. We show that the problem of computing the relative entropy of unambiguous probabilistic automata can be formulated as a shortest-distance problem over an appropriate semiring, give efficient exact and approximate algorithms for its computation in that case, and report the results of experiments demonstrating the practicality of our algorithms for very large weighted automata. We also prove that the computation of the relative entropy of arbitrary probabilistic automata is PSPACE-complete.
The relative entropy is used in a variety of machine learning algorithms and applications to measure the discrepancy of two distributions. We examine the use of the symmetrized relative entropy in machine learning algorithms and show that, contrarily to what is suggested by a number of publications, the symmetrized relative entropy is neither positive definite symmetric nor negative definite symmetric, which limits its use and application in kernel methods. In particular, the convergence of training for learning algorithms is not guaranteed when the symmetrized relative entropy is used directly as a kernel, or as the operand of an exponential as in the case of Gaussian Kernels.
Finally, we show that our algorithm for the computation of the entropy of an unambiguous probabilistic automaton can be generalized to the computation of the norm of an unambiguous probabilistic automaton by using a monoid morphism. In particular, this yields efficient algorithms for the computation of the Lp -norm of a probabilistic automaton.
Title: Domain Decomposition for Less Regular Subdomains: Overlapping Schwarz in Two Dimensions
Author(s): Dohrmann, Clark R.; Klawonn, Axel; Widlund, Olof B.
Abstract:
In the theory of domain decomposition methods, it is often assumed that each subdomain is the union of a small set of coarse triangles or tetrahedra. In this study, extensions to the existing theory which accommodates subdomains with much less regular shape are presented; the subdomains are only required to be John domains. Attention is focused on overlapping Schwarz preconditioners for problems in two dimensions with a coarse space component of the preconditioner which allows for good results even for coefficients which vary considerably. It is shown that the condition number of the domain decomposition method is bounded by C(1 + H/δ)(1 + log(H/h))
^{2}
, where the constant C independent of the number of subdomains and possible jumps in coefficients between subdomains. Numerical examples are provided which confirm the theory and demonstrate very good performance of the method for a variety of subregions including those obtained when a mesh partitioner is used for the domain decomposition.
Title: Democratizing Content Distribution
Candidate: Freedman, Michael
Advisor(s): Mazieres, David
Abstract:
In order to reach their large audiences, today's Internet publishers primarily use content distribution networks (CDNs) to deliver content. Yet the architectures of the prevalent commercial systems are tightly bound to centralized control, static deployments, and trusted infrastructure, inherently limiting their scope and scale to ensure cost recovery.
To move beyond such shortcomings, this thesis contributes a number of techniques that realize cooperative content distribution. By federating large numbers of unreliable or untrusted hosts, we can satisfy the demand for content by leveraging all available resources. We propose novel algorithms and architectures for three central mechanisms of CDNs: content discovery (where are nearby copies of the client's desired resource?), server selection (which node should a client use?), and secure content transmission (how should a client download content efficiently and securely from its multiple potential sources?).
These mechanisms have been implemented, deployed, and tested in production systems that have provided open content distribution services for more than three years. Every day, these systems answer tens of millions of client requests, serving terabytes of data to more than a million people.
This thesis presents five systems related to content distribution. First, Coral provides a distributed key-value index that enables content lookups to occur efficiently and returns references to nearby cached objects whenever possible, while still preventing any load imbalances from forming. Second, CoralCDN demonstrates how to construct a self-organizing CDN for web content out of unreliable nodes, providing robust behavior in the face of failures. Third, OASIS provides a general-purpose, flexible anycast infrastructure, with which clients can locate nearby or unloaded instances of participating distributed systems. Fourth, as a more clean-slate design that can leverage untrusted participants, Shark offers a distributed file system that supports secure block-based file discovery and distribution. Finally, our authentication code protocol enables the integrity verification of large files on-the-fly when using erasure codes for efficient data dissemination.
Taken together, this thesis provides a novel set of tools for building highly-scalable, efficient, and secure content distribution systems. By enabling the automated replication of data based on its popularity, we can make desired content available and accessible to everybody. And in effect, democratize content distribution.
Title: Typical: Taking the Tedium Out of Typing
Author(s): Grimm, Robert; Harris, Laune; Le, Anh
Abstract:
The implementation of real-world type checkers requires a non-trivial engineering effort. The resulting code easily comprises thousands of lines, which increases the probability of software defects in a component critical to compiler correctness. To make type checkers easier to implement and extend, this paper presents Typical, a domain-specific language and compiler that directly and concisely captures the structure of type systems. Our language builds on the functional core for ML to represent syntax trees and types as variants and to traverse them with pattern matches. It then adds declarative constructs for common type checker concerns, such as scoping rules, namespaces, and constraints on types. It also integrates error checking and reporting with other constructs to promote comprehensive error management. We have validated our system with two real-world type checkers written in Typical, one for Typical itself and the other for C.
Title: Declarative Syntax Tree Engineering* Or, One Grammar to Rule Them All
Author(s): Grimm, Robert
Abstract:
Grammars for many parser generators not only specify a language's syntax but also the corresponding syntax tree. Unfortunately, most parser generators pick a somewhat arbitrary combination of features from the design space for syntax trees and thus lock in specific trade-offs between expressivity, safety, and performance. This paper discusses the three major axes of the design space---specification within or outside a grammar, concrete or abstract syntax trees, and dynamically or statically typed trees---and their impact. It then presents algorithms for automatically realizing all major choices from the same, unmodified grammar with inline syntax tree declarations. In particular, this paper shows how to automatically (1) extract a separate syntax tree specification, (2) embed an abstract syntax tree within a concrete one, and (3) infer a strongly typed view on a dynamically typed tree. All techniques are implemented in the Rats! parser generator and have been applied to real-world C and Java grammars and their syntax trees.
Title: Joint Inference for Information Extraction and Translation
Candidate: Ji, Heng
Advisor(s): Grishman, Ralph
Abstract:
The traditional natural language processing pipeline incorporates multiple stages of linguistic analysis. Although errors are typically compounded through the pipeline, it is possible to reduce the errors in one stage by harnessing the results of the other stages.
This thesis presents a new framework based on component interactions to approach this goal. The new framework applies all stages in a suitable order, with each stage generating multiple hypotheses and propagating them through the whole pipeline. Then the feedback from subsequent stages is used to enhance the target stage by re-ranking these hypotheses, and then produce the best analysis.
The effectiveness of this framework has been demonstrated by substantially improving the performance of Chinese and English entity extraction and Chinese-to-English entity translation. The inference knowledge includes mono-lingual interactions among information extraction stages such as name tagging, coreference resolution, relation extraction and event extraction, as well as cross-lingual interaction between information extraction and machine translation.
Such symbiosis of analysis components allows us to incorporate information from a much wider context, spanning the entire document and even going across documents, and utilize deeper semantic analysis; it will therefore be essential for the creation of a high- performance NLP pipeline.
Title: An analysis of a FETI--DP algorithm on irregular subdomains in the plane
Author(s): Klawonn, Axel; Rheinbach, Oliver; Widlund, Olof B.
Abstract:
In the theory for domain decomposition algorithms of the iterative substructuring family, each subdomain is typically assumed to be the union of a few coarse triangles or tetrahedra. This is an unrealistic assumption, in particular, if the subdomains result from the use of a mesh partitioner in which case they might not even have uniformly Lipschitz continuous boundaries.
The purpose of this study is to derive bounds for the condition number of these preconditioned conjugate gradient methods which depend only on a parameter in an isoperimetric inequality and two geometric parameters characterizing John and uniform domains. A related purpose is to explore to what extent well known technical tools previously developed for quite regular subdomains can be extended to much more irregular subdomains.
Some of these results are valid for any John domains, while an extension theorem, which is needed in this study, requires that the subdomains are uniform. The results, so far, are only complete for problems in two dimensions. Details are worked out for a FETI--DP algorithm and numerical results support the findings. Some of the numerical experiments illustrate that care must be taken when selecting the scaling of the preconditioners in the case of irregular subdomains.
Title: Degeneracy Proof Predicates for the Additively Weighted Voronoi Diagram
Candidate: Millman, David
Advisor(s): Yap, Chee
Abstract:
This thesis focuses on the Additively Weighted Voronoi diagram. It begins by presenting the history of the diagram and some of the early algorithms used for its generation [OBSC00, Aur91]. The paper then addresses the more recent incremental approach to calculating the diagram, as seen in the 2D Apollonius Graphs (Delaunay Graphs of Disks) package of CGAL [KY06]. Next, the algorithm of Boissonnat et al. [BD05] for calculating Convex Hulls is presented. We then apply the predicates presented by Bossonnat to the CGAL implementation of the AW-Voronoi diagram, and the results are discussed. The main contribution of this paper results in predicates of the AW-Voronoi diagram, with a lower algebraic degree which also handle degeneracies in such a way as to produce a conical result.
Title: Cellstorm: A bioinformatics software system to visualize subcellular networks
Candidate: Neves, Ana
Advisor(s): Shasha, Dennis
Abstract:
Cellstorm is a software system that allows a rapid visualization of genes and subcellular networks. Given a set of genes, expression levels, structural hierarchy and network's data, Cellstorm displays the networks at any level of the hierarchy and provides a set of user options such as zooming, network selection and list filtering.
Although Cellstorm is mainly aimed at biological applications, it can be used in any field that needs to display networks. Cellstorm achieves this by avoiding application-specific assumptions.
Title: Authentication Mechanisms for Open Distributed Systems
Candidate: Nicolosi, Antonio
Advisor(s): Mazieres, David; Shoup, Victor
Abstract:
While authentication within organizations is a well-understood problem, traditional solutions are often inadequate at the scale of the Internet, where the lack of a central authority, the open nature of the systems, and issues such as privacy and anonymity create new challenges. For example, users typically establish dozens of web accounts with independently administered services under a single password, which increases the likelihood of exposure of their credentials; users wish to receive email from anyone who is not a spammer, but the openness of the email infrastructure makes it hard to authenticate legitimate senders; users may have a rightful expectation of privacy when viewing widely-accessed protected resources such as premium website content, yet they are commonly required to present identifying login credentials, which permits tracking of their access patterns.
This dissertation describes enhanced authentication mechanisms to tackle the challenges of each of the above settings. Specifically, the dissertation develops: 1) a remote authentication architecture that lets users recover easily in case of password compromise; 2) a social network-based email system in which users can authenticate themselves as trusted senders without disclosing all their social contacts; and 3) a group access-control scheme where requests can be monitored while affording a degree of anonymity to the group member performing the request.
The proposed constructions combine system designs and novel cryptographic techniques to address their respective security and privacy requirements both effectively and efficiently.
Title: New Design Criteria for Hash Functions and Block Ciphers
Candidate: Puniya, Prashant
Advisor(s): Dodis, Yevgeniy
Abstract:
Cryptographic primitives, such as hash functions and block ciphers, are integral components in several practical cryptographic schemes. In order to prove security of these schemes, a variety of security assumptions are made on the underlying hash function or block cipher, such as collision-resistance, pseudorandomness etc. In fact, such assumptions are often made without much regard for the actual constructions of these primitives. In this thesis, we address this problem and suggest new, and possibly better, design criteria for hash functions and block ciphers.
We start by analyzing the design criteria underlying hash functions. The usual design principle here involves a two-step procedure: First, come up with a heuristically-designed and ``hopefully strong'' fixed-length input construction (i.e. the compression function), then use a standard domain extension technique, usually the cascade construction, to get a construction that works for variable-length inputs. We investigate this design principle from two perspectives:
We next move on to discuss the Feistel network, which is used in the design of several popular block ciphers such as DES, Triple-DES, Blowfish etc. Currently, the celebrated result of Luby-Rackoff (and further extensions) is regarded as the theoretical basis for using this construction in block cipher design, where it was shown that a four-round Feistel network is a (strong) pseudorandom permutation (PRP) if the round functions are independent pseudorandom functions (PRFs). We study the Feistel network from two different perspectives:
We give a positive answer to the first question and a partial positive answer to the second question. In the process, we undertake a combinatorial study of the Feistel network, that might be useful in other scenarios as well. We provide several practical applications of our results for the Feistel network.
Title: Empirical Bayes least squares estimation without an explicit prior
Author(s): Raphan, Martin; Simoncelli, Eero P.
Abstract:
Bayesian estimators are commonly constructed using an explicit prior model. In many applications, one does not have such a model, and it is difficult to learn since one does not have access to uncorrupted measurements of the variable being estimated. In many cases however, including the case of contamination with additive Gaussian noise, the Bayesian least squares estimator can be formulated directly in terms of the distribution of noisy measurements. We demonstrate the use of this formulation in removing noise from photographic images. We use a local approximation of the noisy measurement distribution by exponentials over adaptively chosen intervals, and derive an estimator from this approximate distribution. We demonstrate through simulations that this adaptive Bayesian estimator performs as well or better than previously published estimators based on simple prior models.
Title: DNA Hash Pooling and its Applications
Author(s): Shasha, Dennis; Amos, Martyn
Abstract:
In this paper we describe a new technique for the characterisation of populations of DNA strands. Such tools are vital to the study of ecological systems, at both the micro (e.g., individual humans) and macro (e.g., lakes) scales. Existing methods make extensive use of DNA sequencing and cloning, which can prove costly and time consuming. The overall objective is to address questions such as: (i) (Genome detection) Is a known genome sequence present at least in part in an environmental sample? (ii) (Sequence query) Is a specific fragment sequence present in a sample? (iii) (Similarity Discovery) How similar in terms of sequence content are two unsequenced samples?
We propose a method involving multiple filtering criteria that result in ``pools" of DNA of high or very high purity. Because our method is similar in spirit to hashing in computer science, we call the method {\it DNA hash pooling}. To illustrate this method, we describe examples using pairs of restriction enzymes. The {\it in silico} empirical results we present reflect a sensitivity to experimental error. The method requires minimal DNA sequencing and, when sequencing is required, little or no cloning.
Title: Being Lazy and Preemptive at Learning toward Information Extraction
Candidate: Shinyama, Yusuke
Advisor(s): Sekine, Satoshi
Abstract:
This thesis proposes a novel approach for exploring Information Extraction scenarios. Information Extraction, or IE, is a task aiming at finding events and relations in natural language texts that meet a user's demand. However, it is often difficult to formulate, or even define such events that satisfy both a user's need and technical feasibility. Furthermore, most existing IE systems need to be tuned for a new scenario with proper training data in advance. So a system designer usually needs to understand what a user wants to know in order to maximize the system performance, while the user has to understand how the system will perform in order to maximize his/her satisfaction.
In this thesis, we focus on maximizing the variety of scenarios that the system can handle instead of trying to improve the accuracy of a particular scenario. In traditional IE systems, a relation is defined a priori by a user and is identified by a set of patterns that are manually crafted or acquired in advance. We propose a technique called Unrestricted Relation Discovery, which defers determining what is a relation and what is not until the very end of the processing so that a relation can be defined a posteriori. This laziness gives huge flexibility to the types of relations the system can handle. Furthermore, we use the notion of recurrent relations to measure how useful each relation is. This way, we can discover new IE scenarios without fully specifying definitions or patterns, which leads to Preemptive Information Extraction, where a system can provide a user a portfolio of extractable relations and let the user choose them.
We used one year news articles obtained from the Web as a development set. We discovered dozens of scenarios that are similar to the existing scenarios tried by many IE systems, as well as new scenarios that are relatively novel. We have evaluated the existing scenarios with Automatic Content Extraction (ACE) event corpus and obtained reasonable performance. We believe this system will shed new light on IE research by giving various experimental IE scenarios.
Title: Constituent Parsing by Classification
Candidate: Turian, Joseph
Advisor(s): Melamed, I. Dan
Abstract:
We present an approach to constituent parsing, which is driven by classifiers induced to minimize a single regularized objective. It is the first discriminatively-trained constituent parser to surpass the Collins (2003) parser without using a generative model. Our primary contribution is simplifying the human effort required for feature engineering. Our model can incorporate arbitrary features of the input and parse state. Feature selection and feature construction occur automatically, as part of learning. We define a set of fine-grained atomic features, and let the learner induce informative compound features. Our learning approach includes several novel approximations and optimizations which improve the efficiency of discriminative training. We introduce greedy completion, a new agenda-driven search strategy designed to find low-cost solutions given a limit on search effort. The inference evaluation function was learned accurately enough to guide the deterministic parsers to the optimal parse reasonably quickly without pruning, and thus without search errors. Experiments demonstrate the flexibility of our approach, which has also been applied to machine translation (Wellington et. al, AMTA 2006; Turian et al., NIPS 2007).
Title: Enhanced Security Models for Network Protocols
Candidate: Walfish, Shabsi
Advisor(s): Dodis, Yevgeniy
Abstract:
Modeling security for protocols running in the complex network environment of the Internet can be a daunting task. Ideally, a security model for the Internet should provide the following guarantee: a protocol that "securely" implements a particular task specification will retain all the same security properties as the specification itself, even when an arbitrary set of protocols runs concurrently on the same network. This guarantee must hold even when other protocols are maliciously designed to interact badly with the analyzed protocol, and even when the analyzed protocol is composed with other protocols. The popular Universal Composability (UC) security framework aims to provide this guarantee.
Unfortunately, such strong security guarantees come with a price: they are impossible to achieve without the use of some trusted setup. Typically, this trusted setup is global in nature, and takes the form of a Public Key Infrastructure (PKI) and/or a Common Reference String (CRS). However, the current approach to modeling security in the presence of such setups falls short of providing expected security guarantees. A quintessential example of this phenomenon is the deniability concern: there exist natural protocols that meet the strongest known security notions (including UC) while failing to provide the same deniability guarantees that their task specifications imply they should provide.
We introduce the Generalized Universal Composability (GUC) framework to extend the UC security notion and enable the re-establishment of its original intuitive security guarantees even for protocols that use global trusted setups. In particular, GUC enables us to guarantee that secure protocols will provide the same level of deniability as the task specification they implement. To demonstrate the usefulness of the GUC framework, we first apply it to the analysis and construction of deniable authentication protocols. Building upon such deniable authentication protocols, we then prove a general feasibility result showing how to construct protocols satisfying our security notion for a large class of two-party and multi-party tasks (assuming the availability of some reasonable trusted setup). Finally, we highlight the practical applicability of GUC by constructing efficient protocols that securely instantiate two common cryptographic tasks: commitments and zero-knowledge proofs.
Title: Tree-Structured Models of Multitext: Theory, Design and Experiments
Candidate: Wellington, Benjamin
Advisor(s): Melamed, I. Dan
Abstract:
Statistical machine translation (SMT) systems use empirical models to simulate the act of human translation between language pairs. This dissertation surveys the ability of currently popular syntax-aware SMT systems to model real-world multitext, and shows different types of linguistic phenomena occurring in natural language translation that these popular systems cannot capture. It then proposes a new grammar formalism, Generalized Multitext Grammar (GMTG), and a generalization of Chomsky Normal Form, that allows us to build an efficient SMT system using previously developed parsing techniques. The dissertation addresses many software engineering issues that arise when doing syntax-based SMT using large corpora and lays out a object-oriented design for a translation toolkit. Using the toolkit, we show that a tree-transduction based SMT system, which uses modern machine learning algorithms, outperforms a generative baseline.
Title: Formal Verification Using Static and Dynamic Analyses
Candidate: Zaks, Aleksandr
Advisor(s): Pnueli, Amir
Abstract:
One of the main challenges of formal verification is the ability to handle systems of realistic size, which is especially exacerbated in the context of software verification. In this dissertation, we suggest two related approaches that, while both rely on formal method techniques, they can still be applied to larger practical systems. The scalability is mainly achieved by restricting the types of properties we are considering and guarantees that are given.
Our first approach is a novel run-time monitoring framework. Unlike previous work on this topic, we expect the properties to be specified using Property Specification Language (PSL). PSL is a newly adopted IEEE P1850 standard and is an extension of Linear Temporal Logic (LTL). The new features include regular expressions and finite trace semantics, which make the new logic very attractive for run-time monitoring of both software and hardware designs. To facilitate the new logic we have extended the existing algorithm for LTL tester construction to cover the PSL specific operators. Another novelty of our approach is the ability to use partial information about the program that is being monitored while the existing tools only use the information about the observed trace and the property under consideration. This allows going beyond the focus of traditional run-time monitoring tools -- error detection in the execution trace, towards the focus of static analysis -- bug detection in programs.
In our second approach, we employ static analysis to compute SAT-based function summaries to detect invalid pointer accesses. To compute function summaries, we propose new techniques for improving the precision and performance in order to reduce the false error rates. In particular, we use BDDs to represent a symbolic simulation of functions, where BDDs allow an efficient representation of path-sensitive information and high level simplification. In addition, we use light-weight range analysis technique for determining lower and upper bounds for program variables, which can further offload the work form the SAT solver. Note that while in our current implementation the analysis happens at compile time, we can also use the function summaries as a basis for run-time monitoring.
Title: A Unified Construction of the Glushkov, Follow, and Antimirov Automata
Author(s): Allauzen, Cyril; Mohri, Mehryar
Abstract:
Many techniques have been introduced in the last few decades to create ε-free automata representing regular expressions: Glushkov automata, the so-called follow automata, and Antimirov automata. This paper presents a simple and unified view of all these ε-free automata both in the case of unweighted and weighted regular expressions.It describes simple and general algorithms with running time complexities at least as good as that of the best previously known techniques, and provides concise proofs.The construction methods are all based on two standard automata algorithms: epsilon-removal and minimization. This contrasts with the multitude of complicated and special-purpose techniques and proofs put forward by others to construct these automata. Our analysis provides a better understanding of ε-free automata representing regular expressions: they are all the results of the application of some combinations of epsilon-removal and minimization to the classical Thompson automata. This makes it straight forward to generalize these algorithms to the weighted case, which also results in much simpler algorithms than existing ones. For weighted regular expressions over a closed semiring, we extend the notion of follow automata to the weighted case. We also present the first algorithm to compute the Antimirov automata in the weighted case.
Title: Shape Analysis of Single-Parent Heaps
Author(s): Balaban, Ittai; Pnueli, Amir; Zuck, Lenore
Abstract:
We define the class of single-parent heap systems, which rely on a singly-linked heap in order to model destructive updates on tree structures. This encoding has the advantage of relying on a relatively simple theory of linked lists in order to support abstraction computation. To facilitate the application of this encoding, we provide a program transformation that, given a program operating on a multi-linked heap without sharing, transforms it into one over a single-parent heap. It is then possible to apply shape analysis by predicate and ranking abstraction as in [BPZ05]. The technique has been successfully applied on examples with trees of fixed arity (balancing of and insertion into a binary sort tree).
Title: Invisible Safety of Distributed Protocols
Author(s): Balaban, Ittai; Pnueli, Amir; Zuck, Lenore
Abstract:
The method of ``Invisible Invariants'' has been applied successfully to protocols that assume a ``symmetric'' underlying topology, be it cliques, stars, or rings. In this paper we show how the method can be applied to proving safety properties of distributed protocols running under arbitrary topologies. Many safety properties of such protocols have reachability predicates, which, on first glance, are beyond the scope of the Invisible Invariants method. To overcome this difficulty, we present a technique, called ``coloring,'' that allows, in many instances, to replace the second order reachability predicates by first order predicates, resulting in properties that are amenable to Invisible Invariants, where ``reachable'' is replaced by ``colored.'' We demonstrate our techniques on several distributed protocols, including a variant on Luby's Maximal Independent Set protocol, the Leader Election protocol used in the IEEE 1394 (Firewire) distributed bus protocol, and various distributed spanning tree algorithms. All examples have been tested using the symbolic model checker TLV.
Title: TimeIn: A temporal visualization for file access
Candidate: Borden, Jeffrey
Advisor(s): Shasha, Dennis
Abstract:
TimeIn seeks to unify a given set of file objects into a unified browsing experience providing mechanisms to Cluster visually similar objects and display objects in a timeline view from a local file system or flickr.com . To navigate this content, users are provided with a variety of mechanisms for filtering the set of objects presented.
For text based objects, TimeIn will analyze the content of the file and attempt to extract a set of descriptive phrases. For image based objects, TimeIn will annotate the object with the most frequently used colors of the image. Users have the option of augmenting these automatically generated tags by defining their own descriptive tags.
While providing novel features for browsing and searching content, TimeIn retains many of the original organizational features of existing systems. When content is imported from a hierarchical file system, users can still browse by the original hierarchical structure.
TimeIn also retains the PhotoSet structures associated with content imported from flickr.com . Users can also organize content into user-defined "albums" of objects. These albums can then be used to filter the set of objects on the timeline.
Title: On Transductive Regression
Author(s): Cortes, Corinna; Mohri, Mehryar
Abstract:
In many modern large-scale learning applications, the amount of unlabeled data far exceeds that of labeled data. A common instance of this problem is the 'transductive' setting where the unlabeled test points are known to the learning algorithm. This paper presents a study of regression problems in that setting. It presents 'explicit' VC-dimension error bounds for transductive regression that hold for all bounded loss functions and coincide with the tight classification bounds of Vapnik when applied to classification. It also presents a new transductive regression algorithm inspired by our bound that admits a primal and kernelized closed-form solution and deals efficiently with large amounts of unlabeled data. The algorithm exploits the position of unlabeled points to locally estimate their labels and then uses a global optimization to ensure robust predictions. Our study also includes the results of experiments with several publicly available regression data sets with up to 20,000 unlabeled examples. The comparison with other transductive regression algorithms shows that it performs well and that it can scale to large data sets.
Title: Guaranteed Precision for Transcendental and Algebraic Computation Made Easy
Candidate: Du, Zilin
Advisor(s): Yap, Chee
Abstract:
Numerical non-robustness is a well-known phenomenon when implementing geometric algorithms. A general approach to achieve geometric robustness is Exact Geometric Computation (EGC). This dissertation explores the redesign and extension of Core Library, a C++ library which embraces the EGC approach. The contributions of this thesis are organized into three parts.
In the first part, we discuss the redesign of Core Library, especially the expression "Expr" and bigfloat "BigFloat" classes. Our new design emphasizes extensibility in a clean and modular way. The three facilities in "Expr", filter, root bound and bigfloat, are separated into independent modules. This allows new filters, root bounds and some bigfloat substitute to be plugged in. The key approximate evaluation and precision propagation algorithms have been greatly improved. A new bigfloat system based on MPFR and interval arithmetic has been incorporated. Our benchmark shows that the redesigned Core Library typically has 5-10 times speedup. We also provide tools to facilitate extensions of "Expr" to incorporate new type of nodes, especially transcendental nodes.
Although the Core Library was originally designed for algebraic applications, transcendental functions are needed in many applications. In the second part, we present a complete algorithm for absolute approximation of the general hypergeometric functions. It's complexity is also given. The extension of this algorithm to ``blackbox number'' is provided. A general hypergeometric function package based on our algorithm is implemented and integrated into the Core Library based on our new design.
Brent has shown that many elementary functions, such as $\exp, \log, \sin$, etc., can be efficiently computed using the Arithmetic-Geometric Mean (AGM) based algorithm. However, he only gave an asymptotic error analysis. The constants in the Big $O(\cdot)$ notation required for implementation are unknown. We provide a non-asymptotic error analysis of the AGM algorithm and the related algorithms for logarithm and exponential functions. These algorithms have been implemented and incorporated into the Core Library.
Title: On Cryptographic Techniques for Digital Rights Management
Candidate: Fazio, Nelly
Advisor(s): Dodis, Yevgeniy
Abstract:
With more and more content being produced, distributed, and ultimately rendered and consumed in digital form, devising effective Content Protection mechanisms and building satisfactory Digital Rights Management (DRM) systems have become top priorities for the Publishing and Entertaining Industries.
To help tackle this challenge, several cryptographic primitives and constructions have been proposed, including mechanisms to securely distribute data over a unidirectional insecure channel (Broadcast Encryption), schemes in which leakage of cryptographic keys can be traced back to the leaker (Traitor Tracing), and techniques to combine revocation and tracing capabilities (Trace-and-Revoke schemes).
In this thesis, we present several original constructions of the above primitives, which improve upon existing DRM-enabling cryptographic primitives along the following two directions:
Our results along the first line of work include the following:
As for the second direction, our contribution can be divided as follows:
Overall, the cryptographic tools developed in this thesis provide more flexibility and more security than existing solutions, and thus offer a better match for the challenges of the DRM setting.
Title: Finding Your Match: Techniques for Improving Sequence Alignment in DNA and RNA
Candidate: Gill, Ofer Hirsch
Advisor(s): Mishra, Bud
Abstract:
In Bioinformatics, finding correlations between species allows us the better understand the important biological functions of those species and trace its evolution. This thesis considers sequence alignment, a method for obtaining these correlations. We improve upon sequence alignment tools designed for DNA with Plains, an algorithm than uses piecewise-linear gap functions and parameter-optimization to obtain correlations in remotely-related species pairs such as human and fugu using reasonable amounts of memory and space on an ordinary computer. We then discuss Planar, which is similar to Plains, but is designed for aligning RNA, and accounts for secondary structure. We also explore SEPA, a tool that uses p-value estimation based on exhaustive empirical data to better emphasize key results from an alignment with a measure of reliability. Using SEPA to measure the quality of an alignment, we proceed to compare Plains and Planar against similar alignment tools, emphaisizing the interesting correlations caught in the process.
Title: Kronosphere: A temporal visualization for file access
Candidate: Harrison, Chris
Advisor(s): Shasha, Dennis
Abstract:
Hierarchical file systems mirror the way people organize data in the real world. However, this method of organization is often inadequate in managing the immense number of files that populate hard drives. Kronosphere provides a novel time and content-based navigational paradigm for managing and accessing media. This allows users to browse their documents by time, content, history, metadata, and relationships with other files.
Title: DataSlicer: A Hosting Platform for Data-Centric Network Services
Candidate: He, Congchun
Advisor(s): Karamcheti, Vijay
Abstract:
As the Web evolves, the number of network services deployed on the Internet has been growing at a dramatic pace. Such services usually involve a massive volume of data stored in physical or virtual back-end databases, and access the data to dynamically generate responses for client requests. These characteristics restrict use of traditional mechanisms for improving service performance and scalability: large volumes prevent replication of the service data at multiple sites required by content distribution schemes, while dynamic responses do not support the reuse required by web caching schemes.
However, many deployed data-centric network services share other properties that can help alleviate this situation: (1) service usage patterns exhibit locality of various forms, and (2) services are accessed using standard protocols and publicly known message structures. When properly exploited, these characteristics enable the design of alternative caching infrastructures, which leverage distributed network intermediaries to inspect traffic flowing between clients and services, infer locality information dynamically, and potentially improve service performance by taking actions such as partial service replication, request redirection, or admission control.
This dissertation investigates the nature of locality in service usage patterns for two well-known web services, and reports on the design, implementation, and evaluation of such a network intermediary architecture, named DataSlicer. DataSlicer incorporates four main techniques: (1) Service-neutral request inspection and locality detection on distributed network intermediaries; (2) Construction of oriented overlays for clustering client requests; (3)Integrated load-balancing and service replication mechanisms that improve service performance and scalability by either redistributing the underlying traffic in the network or creating partial service replicas on demand at appropriate network locations; and (4) Robustness mechanisms to maintain system stability in a wide-area network environment.
DataSlicer has been successfully deployed on the PlanetLab network. Extensive experiments using synthetic workloads show that our approach can: (1) create appropriate oriented overlays to cluster client requests according to multiple application metrics; (2) detect locality information across multiple dimensions and granularity levels; (3) leverage the detected locality information to perform appropriate load-balancing and service replication actions with minimal cost; and (4) ensure robust behavior in the face of dynamically changing network conditions.
Title: Multimarker Genetic Analysis Methods for High Throughput Array Data
Candidate: Ionita, Iuliana
Advisor(s): Mishra, Bud
Abstract:
In this thesis, we focus on multi-marker/-locus statistical methods for analyzing high-throughput array data used for the detection of genes implicated in complex disorders. There are two main parts: the first part concerns the localization of cancer genes from copy number variation data, with an application to lung cancer; the second part concerns the localization of disease genes using an affected-sib-pair design, with an application to inflammatory bowel disease. A third part addresses an important issue involved in the design of these disease-gene-detection studies. More details follow:
1. Detection of Oncogenes and Tumor Suppressor Genes using Multipoint Statistics from Copy Number Variation Data
ArrayCGH is a microarray-based comparative genomic hybridization technique that has been used to compare a tumor genome against a normal genome, thus providing rapid genomic assays of tumor genomes in terms of copy number variations of those chromosomal segments, which have been gained or lost. When properly interpreted, these assays are likely to shed important light on genes and mechanisms involved in initiation and progression of cancer. Specifically, chromosomal segments, amplified or deleted in a group of cancer patients, point to locations of cancer genes. We describe a statistical method to estimate the location of such genes by analyzing segmental amplifications and deletions in the genomes from cancer patients and the spatial relation of these segments to any specific genomic interval. The algorithm assigns to a genomic segment a score that parsimoniously captures the underlying biology. It computes a p-value for every putative disease gene by using results from the theory of scan statistics. We have validated our method using simulated datasets, as well as a real dataset on lung cancer.
2. Multi-locus Linkage Analysis of Affected-Sib-Pairs
A The affected-sib-pair (ASP) design is a simple and popular design in the linkage analysis of complex traits. The traditional ASP methods evaluate the linkage information at a locus by considering only the marginal linkage information present at that locus. However complex traits are influenced by multiple genes that together interact to increase the risk to disease. We describe a multi-locus linkage method that uses both the marginal information and information derived from the possible interactions among several disease loci, thereby increasing the significance of loci with modest marginal effects. Our method is based on a statistic that quantifies the linkage information contained in a set of markers. By a marker selection-reduction process, we screen a set of polymorphisms and select a few that seem linked to disease. We test our approach on simulated data and a genome-scan data for inflammatory bowel disease. We show that our method is expected to be more powerful than single-locus methods in detecting disease loci responsible for complex traits.
3. A Practical Haplotype Inference Algorithm
We consider the problem of efficient inference algorithms to determine the haplotypes and their distribution from a dataset of unrelated genotypes.
With the currently available catalogue of single-nucleotide polymorphisms (SNPs) and given their abundance throughout the genome (one in about $500$ bps) and low mutation rates, scientists hope to significantly improve their ability to discover genetic variants associated with a particular complex trait. We present a solution to a key intermediate step by devising a practical algorithm that has the ability to infer the haplotype variants for a particular individual from its own genotype SNP data in relation to population data. The algorithm we present is simple to describe and implement; it makes no assumption such as perfect phylogeny or the availability of parental genomes (as in trio-studies); it exploits locality in linkages and low diversity in haplotype blocks to achieve a linear time complexity in the number of markers; it combines many of the advantageous properties and concepts of other existing statistical algorithms for this problem; and finally, it outperforms competing algorithms in computational complexity and accuracy, as demonstrated by the studies performed on real data and synthetic data.
Title: A FETI-DP algorithm for elasticity problems with mortar discretization on geometrically non-conforming partitions
Author(s): Kim, Hyea Hyun
Abstract:
In this paper, a FETI-DP formulation for three dimensional elasticity on non-matching grids over geometrically non-conforming subdomain partitions is considered. To resolve the nonconformity of the finite elements, a mortar matching condition is imposed on the subdomain interfaces (faces). A FETI-DP algorithm is then built by enforcing the mortar matching condition in dual and primal ways. In order to make the FETI-DP algorithm scalable, a set of primal constraints, which include average and momentum constraints over interfaces, are selected from the mortar matching condition. A condition number bound, $C(1+\text{log}(H/h))2$, is then proved for the FETI-DP formulation for the elasticity problems with discontinuous material parameters. Only some faces need to be chosen as primal faces on which the average and momentum constraints are imposed.
Title: Expressive Motion
Candidate: Lees, Alyssa
Advisor(s): Bregler, Christopher; Geiger, Davi
Abstract:
Since the advent of motion capture animation, attempts have been made to extract the seemingly nebulously defined attributes of 'content' and 'style' from the motion data. Enabling quick access to highly precise data, the benefits of motion capture for animation purposes are abundant. Yet manipulating the expressive attributes of the motion data in a comprehensive manner has proved elusive. This dissertation poses practical solutions that are based on insights from the dance community and learning attributes from the motion data itself. The culminating project is a system which learns the deformations of the human body and reapplies them in exaggerated form for enhanced expressivity.
While simultaneously developing efficient and usable tools for animators, the result is a three pronged technique to enhance the expressive qualities of motion capture animation. The key aspect is the creation of a deformable skeleton representation of the human body using a unique machine learning approach. The deformable skeleton is modeled by replicating the actual movements of the human spine. The second step relies on exploiting the subtle aspects of motion, such as hand movement to create an emotional effect visually. Both of these approaches involve exaggerating the movements in the same vein as traditional 2-D animation technique of 'squash and stretch'. Finally, a novel technique for the application of style on a baseline motion capture sequence is developed.
All of these approaches are rooted in machine learning techniques. Linear discriminate analysis was initially applied to a single phrase of motion demonstrating various style characteristics in LABAN notation. A variety of methods including nonlinear PCA, and LLE were used to learn the underlying manifold of spine movements. Nonlinear dynamic models were learned in attempts to describe motion segments versus single phrases. In addition, the dissertation focuses on the variety of obstacles in learning with motion data. This includes the correct parameterization of angles, applying statistical analysis to quaternions, and appropriate distance measures between postures.
Title: Building Trustworthy Storage Services out of Untrusted Infrastructure
Candidate: Li, Jinyuan
Advisor(s): Mazieres, David
Abstract:
As the Internet has become increasingly ubiquitous, it has seen tremendous growth in the popularity of online services. These services range from online CVS repositories like sourceforge , shopping sites, to online financial and administrative systems, etc. It is critical for these services to provide correct and reliable execution for clients. However, given their attractiveness as targets and ubiquitous accessibility, online servers also have a significant chance of being compromised, leading to Byzantine failures.
Designing and implementing a service to run on a machine that may be compromised is not an easy task, since infrastructure under malicious control may behave arbitrarily. Even worse, as any monitoring facility may also be subverted at the same time, there is no easy way for system behavior to be audited, or for malicious attacks to be detected.
We propose our solution to the problem by reducing the trust needed on the server side in the first place. In the other words, our system is designed specifically for running on untrusted hosts. In this thesis, we realize this principle by two different approaches. First, we design and implement a new network file system -- SUNDR. In SUNDR, malicious servers cannot forge users' operations or tamper with their data without being detected. In the worst case, attackers can only conceal users' operations from each other. Still, SUNDR is able to detect this misbehavior whenever users communicate with each other directly.
The limitation of the approach above lies in that the system cannot guarantee ideal consistency with even one single failure. In the second approach, we use replicated state machines to tolerate some fraction of malicious server failures, which is termed Byzantine Fault Tolerance (BFT) in the literature. Classical BFT systems assume less than 1/3 of the replicas are malicious, to provide ideal consistency. In this thesis, we push the boundary from 1/3 to 2/3. With fewer than 1/3 of replicas faulty, we provide the same guarantees as classical BFT systems. Additionally, we guarantee weaker consistency, instead of arbitrary behavior, when between 1/3 and 1/3 of replicas fail.
Title: Measures for Robust Stability and Controllability
Candidate: Mengi, Emre
Advisor(s): Overton, Michael
Abstract:
A linear time-invariant dynamical system is robustly stable if the system as well as all of its nearby systems in a neighborhood of interest are stable. An important property of robustly stable systems is they decay asymptotically without exhibiting significant transient behavior. The first part of this thesis work focuses on measures revealing the degree of robust stability of a dynamical system. We put special emphasis on pseudospectral measures, those based on the eigenvalues of nearby matrices for a first-order system or matrix polynomials for a higher-order system. We present algorithms for the computation of pseudospectral measures for continuous and discrete systems with quadratic rate of convergence and analyze their accuracy in the presence of rounding errors. We also provide an efficient algorithm for the numerical radius of a matrix, the modulus of the outermost point in the field of values (the set of Rayleigh quotients) of the matrix. These algorithms are inspired by algorithms of Byers, Boyd-Balakrishnan and Burke-Lewis-Overton.
The second part is devoted to indicators of robust controllability. We call a system robustly controllable if it is controllable and remains controllable under perturbations of interest. We describe efficient methods for the computation of the distance to the closest uncontrollable system. Our first algorithm for the first-order distance to uncontrollability depends on a grid and is well-suited for low precision approximation. We then discuss algorithms for high precision approximation of the first-order distance to uncontrollability. These are based on the bisection method of Gu and the trisection variant of Burke-Lewis-Overton.
These algorithms require the extraction of the real eigenvalues of matrices of size $O(n2)$ typically at a cost of $O(n6)$, where $n$ is the dimension of the state space. We propose a new divide-and-conquer algorithm that reduces the cost to $O(n4)$ on average in both theory and practice and $O(n5)$ in the worst case. The new iterative approach to the extraction of real eigenvalues may also be useful in other contexts. For higher-order systems we derive a singular value characterization and exploit this characterization for the computation of the higher-order distance to uncontrollability to low precision. The algorithms in this thesis assume arbitrary complex perturbations are applicable to the input system and usually require the extraction of the imaginary eigenvalues of Hamiltonian matrices (or even matrix polynomials) or the unit eigenvalues of symplectic pencils (or palindromic matrix polynomials).
Title: Algorithmic Algebraic Model Checking: Hybrid Automata & Systems Biology
Candidate: Mysore, Venkatesh Pranesh
Advisor(s): Mishra, Bud
Abstract:
Systems Biology strives to hasten our understanding of the fundamental principles of life by adopting a systems-level approach for the analysis of cellular function and behavior. One popular framework for capturing the chemical kinetics of interacting biochemicals is Hybrid Automata. Our goal in this thesis is to aid Systems Biology research by improving the current understanding of hybrid automata, by developing techniques for symbolic rather than numerical analysis of the dynamics of biochemical networks modeled as hybrid automata, and by honing the theory to two classes of problems: kinetic mass action based simulation in genetic regulatory & signal transduction pathways, and pseudo-equilibrium simulation in metabolic networks.
We first provide new constructions that prove that the "open" Hierarchical Piecewise Constant Derivative (HPCD) subclass is closer to the decidability and undecidability frontiers than was previously understood. After concluding that the HPCD-like classes are unsuitable for modeling chemical reactions, our quest for semi-decidable subclasses leads us to define the "semi-algebraic" subclass. This is the most expressive hybrid automaton subclass amenable to rigorous symbolic temporal reasoning. We begin with the bounded reachability problem, and then show how the dense-time temporal logic Timed Computation Tree Logic (TCTL) can be model-checked by exploiting techniques from real algebraic geometry, primarily real quantifier elimination. We also prove the undecidability of reachability in the Blum-Shub-Smale Turing Machine formalism. We then develop efficient approximation strategies by extending bisimulation partitioning, rectangular grid-based approximation, polytopal approximation and time discretization. We then develop a uniform algebraic framework for modeling biochemical and metabolic networks, also extending flux balance analysis. We present some preliminary results using a prototypical tool Tolque. It is a symbolic algebraic dense time model-checker for semi-algebraic hybrid automata, which uses Qepcad for quantifier elimination.
The "Algorithmic Algebraic Model Checking" techniques developed in this thesis present a theoretically-grounded mathematically-sound platform for powerful symbolic temporal reasoning over biochemical networks and other semi-algebraic hybrid automata. It is our hope that by building upon this thesis, along with the development of computationally efficient parallelizable quantifier elimination algorithms and the integration of different computer algebra tools, scientific software systems will emerge that fundamentally transform the way biochemical networks (and other hybrid automata) are analyzed.
Title: Building an Automatic Phenotyping System of Developing Embryos
Candidate: Ning, Feng
Advisor(s): LeCun, Yann
Abstract:
This dissertation presents a learning-based system for the detection, identification, localization, and measurement of various sub-cellular structures in microscopic images of developing embryos. The system analyzes sequences of images obtained through DIC microscopy and detects cell nuclei, cytoplasm, and cell walls automatically. The system described in this dissertation is the key initial component of a fully automated phenotype analysis system.
Our study primarily concerns the early stages of development of C. Elegans nematode embryos, from fertilization to the four-cell stage. The method proposed in this dissertation consists in learning the entire processing chain {\em from end to end}, from raw pixels to ultimate object categories.
The system contains three modules: (1) a convolutional network trained to classify each pixel into five categories: cell wall, cytoplasm, nuclear membrane, nucleus, outside medium; (2) an Energy-Based Model which cleans up the output of the convolutional network by learning local consistency constraints that must be satisfied by label images; (3) A set of elastic models of the embryo at various stages of development that are matched to the label images.
When observing normal (wild type) embryos it is possible to visualize important cellular functions such as nuclear movements and fusions, cytokinesis and the setting up of crucial cell-cell contacts. These events are highly reproducible from embryo to embryo. The events will deviate from normal behaviors when the function of a specific gene is perturbed, therefore allowing the detection of correlations between genes activities and specific early embryonic events. One important goal of the system is to automatically detect whether the development is normal (and therefore, not particularly interesting), or abnormal and worth investigating. Another important goal is to automatically extract quantitative measurements such as the migration speed of the nuclei and the precise time of cell divisions.
Title: A Polymorphic Type System and Compilation Scheme for Record Concatenation
Candidate: Osinski, Edward
Advisor(s): Goldberg, Benjamin
Abstract:
The notion of records, which are used to organize closely related groups of data so the group can be treated as a unit, and also provide access to the data within by name, is almost universally supported in programming languages. However, in virtually all cases, the operations permitted on records in statically typed languages are extremely limited. Providing greater flexibility in dealing with records, while simultaneously retaining the benefits of static type checking is a desirable goal.
This problem has generated considerable interest, and a number of type systems dealing with records have appeared in the literature. In this work, we present the first polymorphic type system that is expressive enough to type a number of complex operations on records, including three forms of concatenation and natural join. In addition, the precise types of the records involved are inferred, to eliminate the burden of explicit type declarations. Another aspect of this problem is an efficient implementation of records and their associated operations. We also present a compilation method which accomplishes this goal.
Title: A Probabilistic Learning Approach to Attribute Value Inconsistency Resolution
Candidate: Pevzner, Ilya
Advisor(s): Goldberg, Arthur
Abstract:
Resolving inconsistencies in data is a problem of critical practical importance. Inconsistent data arises whenever an attribute takes on multiple, inconsistent, values. This may occur when a particular entity is stored multiple times in one database, or in multiple databases that are combined.
We investigate Attribute Value Inconsistency Resolution (AVIR), the problem of semi-automatically resolving data inconsistencies among multiple database records that describe the same person or thing.
Our survey of the area shows that existing solutions are either limited in scope or impose a significant burden on their users. Either they do not cover all types of inconsistencies and attributes, or they require users to write or choose attribute resolution functions for each potentially conflicting attribute.
Our ML based approach applies to all types of inconsistencies and attributes, and automatically selects appropriate resolution functions based on the conflicting data. We have invented and developed a system, that uses a set of binary features that detect data properties and relationships and resolution functions that merge data. Many such features and resolution functions have been written. The system uses supervised learning with maximum likelihood estimation to determine which function(s) to apply, based on which feature(s) fire.
We have validated our system by comparing its error rate, decision rate and decision accuracy on a test data set to baseline values determined by a clairvoyant application of a standard approach where each potentially conflicting attribute is resolved by the best resolution function for the attribute.
Title: PSL Model Checking and Run-time Verification via Testers
Author(s): Pnueli, Amir; Zaks, Aleksandr
Abstract:
The paper introduces the construct of \emm{temporal testers} as a compositional basis for the construction of automata corresponding to temporal formulas in the PSL logic. Temporal testers can be viewed as (non-deterministic) transducers that, at any point, output a boolean value which is 1 iff the corresponding temporal formula holds starting at the current position.
The main advantage of testers, compared to acceptors (such as Buchi automata) is that they are compositional. Namely, a tester for a compound formula can be constructed out of the testers for its sub-formulas. In this paper, we extend the application of the testers method from LTL to the logic PSL.
Besides providing the construction of testers for PSL, we indicate how the symbolic representation of the testers can be directly utilized for efficient model checking and run-time monitoring.
Title: Animating Autonomous Pedestrians
Candidate: Shao, Wei
Advisor(s): Terzopoulos, Demetri
Abstract:
This thesis addresses the difficult open problem in computer graphics of autonomous human modeling and animation, specifically of emulating the rich complexity of real pedestrians in urban environments.
We pursue an artificial life approach that integrates motor, perceptual, behavioral, and cognitive components within a model of pedestrians as highly capable individuals. Our comprehensive model features innovations in these components, as well as in their combination, yielding results of unprecedented fidelity and complexity for fully autonomous multi-human simulation in large urban environments. Our pedestrian model is entirely autonomous and requires no centralized, global control whatsoever.
To animate a variety of natural interactions between numerous pedestrians and their environment, we represent the environment using hierarchical data structures, which efficiently support the perceptual queries of the autonomous pedestrians that drive their behavioral responses and sustain their ability to plan their actions on local and global scales.
The animation system that we implement using the above models enables us to run long-term simulations of pedestrians in large urban environments without manual intervention. Real-time simulation can be achieved for well over a thousand autonomous pedestrians. With each pedestrian under his/her own autonomous control, the self-animated characters imbue the virtual world with liveliness, social (dis)order, and a realistically complex dynamic.
We demonstrate the automated animation of human activity in a virtual train station, and we employ our pedestrian simulator in the context of virtual archaeology for visualizing urban social life in reconstructed archaeological sites. Our pedestrian simulator is also serving as the basis of a testbed for designing and experimenting with visual sensor networks in the field of computer vision.
Title: Complexity Analysis of Algorithms in Algebraic Computation
Candidate: Sharma, Vikram
Advisor(s): Yap, Chee
Abstract:
Numerical computations with real algebraic numbers require algorithms for approximating and isolating real roots of polynomials. A classical choice for root approximation is Newton's method. For an analytic function on a Banach space, Smale introduced the concept of approximate zeros, i.e., points from which Newton's method for the function converges quadratically. To identify these approximate zeros he gave computationally verifiable convergence criteria called point estimates. However, in developing these results Smale assumed that Newton's method is computed exactly. For a system of $n$ homogeneous polynomials in $n+1$ variables, Malajovich developed point estimates for a different definition of approximate zero, assuming that all operations in Newton's method are computed with fixed precision. In the first half of this dissertation, we develop point estimates for these two different definitions of approximate zeros of an analytic function on a Banach space, but assume the strong bigfloat computational model of Brent, i.e., where all operations involve bigfloats with varying precision. In this model, we derive a uniform complexity bound for approximating a root of a zero-dimensional system of $n$ integer polynomials in $n$ variables. We also derive a non-asymptotic bound, in terms of the condition number of the system, on the precision required to implement the robust Newton method.
The second part of the dissertation analyses the worst-case complexity of two algorithms for isolating real roots of a square-free polynomial with real coefficients: The Descartes method and Akritas' continued fractions algorithm. The analysis of both algorithms is based upon amortization bounds such as the Davenport-Mahler bound. For the Descartes method, we give a unified framework that encompasses both the power basis and the Bernstein basis variant of the method; we derive an $O(n(L+\log n))$ bound on the size of the recursion tree obtained by applying the method to a square-free polynomial of degree n with integer coefficients of bit-length $L$, the bound is tight for $L=\Omega(\log n)$; based upon this result we readily obtain the best known bit-complexity bound of $\wt{O}(n^4L2) $ for the Descartes method, where $\wt{O}$ means we ignore logarithmic factors. Similar worst case bounds on the bit-complexity of Akritas' algorithm were not known in the literature. We provide the first such bound, $\wt{O}(n^{12}L3)$, for a square-free integer polynomial of degree $n$ and coefficients of bit-length $L$.
Title: Pairwise Comparison between Genomic Sequences and Optical-Maps
Candidate: Sun, Bing
Advisor(s): Mishra, Bud
Abstract:
With the development and improvement of high throughput experimental technologies, massive amount of biological data including genomic sequences and optical-maps have been collected for various species. Comparative techniques play a central role in investigating the adaptive significance of organismal traits and revealing evolutionary relations among organisms by comparing these biological data. This dissertation presents two efficient comparative analysis tools used in comparative genomics and comparative optical-map study, respectively.
A complete genome sequence of an organism can be viewed as its ultimate genetic map, in the sense that the heritable information are encoded within the DNA and the order of nucleotides along chromosomes is known. Comparative genomics can be applied to find functional sites by comparing genetic maps. Comparing vertebrate genomes requires efficient cross-species sequence alignment programs. The first tool introduced in this thesis is COMBAT (Clean Ordered Mer-Based Alignment Tool), a new mer-based method which can search rapidly for highly similar translated genomic sequences using the stable-marriage algorithm (SM) as an alignment filter. In experiments COMBAT is applied to comparative analysis between yeast genomes, and between the human genome and the recently published bovine genome. The homologous blocks identified by COMBAT are comparable with the alignments produced by BLASTP and BLASTZ.
When genetic maps are not available, other genomic maps, including optical-maps, can be constructed. An optical map is an ordered enumeration of the restriction sites along with the estimated lengths of the restriction fragments between consecutive restriction sites. CAPO (Comparative Analysis and Phylogeny with Optical-Maps), introduced as a second technique in this thesis, is a tool for inferring phylogeny based on pairwise optical map comparison and bipartite graph matching. CAPO combines the stable matching algorithm with either the Unweighted Pair Group Method with Arithmetic Averaging (UPGMA) or the Neighbor-Joining (NJ) method for constructing phylogenetic trees. This new algorithm is capable of constructing phylogenetic trees in logarithmic steps and performs well in practice. Using optical maps constructed in silico and in vivo, our work shows that both UPGMA-flavored trees and the NJ-flavored trees produced by CAPO share substantial overlapping tree topology and are biologically meaningful.
Title: Exploiting Service Usage Information for Optimizing Server Resource Management
Candidate: Totok, Alexander
Advisor(s): Karamcheti, Vijay
Abstract:
It is difficult to provision and manage modern component-based Internet services so that they provide stable quality-of-service (QoS) guarantees to their clients, because: (1) component middleware are complex software systems that expose several independently tuned configurable application runtime policies and server resource management mechanisms; (2) session-oriented client behavior with complex data access patterns makes it hard to predict what impact tuning these policies and mechanisms has on application behavior; (3) component-based Internet services exhibit complex structural organization with requests of different types accessing different components and data sources, which could be distributed and/or replicated for failover, performance, or business purposes.
This dissertation attempts to alleviate this situation by targeting three interconnected goals: (1) providing improved QoS guarantees to the service clients, (2) optimizing server resource utilization, and (3) providing application developers with guidelines for natural application structuring, which enable efficient use of the proposed mechanisms for improving service performance. Specifically, we explore the thesis that exposing and using detailed information about how clients use component-based Internet services enables mechanisms that achieve the range of goals listed above. To validate this thesis we show its applicability to the following four problems: (1) maximizing reward brought by Internet services, (2) optimizing utilization of server resource pools, (3) providing session data integrity guarantees, and (4) enabling service distribution in wide-area environments.
The techniques that we propose for the identified problems are applicable at both the application structuring stage and the application operation stage, and range from automatic (i.e., performed by middleware in real time) to manual (i.e., involve the programmer, or the service provider). These techniques take into account service usage information exposed at different levels, ranging from high-level structure of user sessions to low level information about data access patterns and resource utilization by requests of different types. To show the benefits of the proposed techniques, we implement various middleware mechanisms in the JBoss application server, which utilizes the J2EE component model, and comprehensively evaluate them on several publicly-available sample J2EE applications - Java Pet Store, RUBiS, and our own implementation of the TPC-W web transactional benchmark. Our experimental results show that the proposed techniques achieve optimal utilization of server resources and improve application performance by up to two times for centralized Internet services and by up to 6 times for distributed ones.
Title: Time Series Matching: A Multi-Filter Approach
Candidate: Wang, Zhihua
Advisor(s): Shasha, Dennis
Abstract:
Data arriving in time order (time series) arises in disciplines ranging from music to meteorology to finance to motion capture data, to name a few. In many cases, a natural way to query the data is what we call time series matching - a user enters a time series by hand, keyboard or voice and the system finds "similar" time series.
Existing time series similarity measures, such as DTW (Dynamic Time Warping), can accommodate certain timing errors in the query and perform with high accuracy on small databases. However, they all have high computational complexity and the accuracy dramatically drops when the data set grows. More importantly, there are types of errors that cannot be captured by a single similarity measure.
Here we present a general time series matching framework. This framework can easily optimize, combine and test different features to execute a fast similarity search based on the application's requirement. Basically we use a multi-filter chain and boosting algorithms to compose a ranking algorithm. Each filter is a classifier which removes bad candidates by comparing certain features of the time series data. Some filters use a boosting algorithm to combine a few different weak classifiers into a strong classifier. The final filter will give a ranked list of candidates in the reference data which matches the query data.
The framework is applied to build query algorithms for a Query-by-Humming system. Experiments show that the algorithm has a more accurate similarity measure and its response time increases much slower than the pure DTW algorithm when the number of songs in the database increases from 60 to 1400.
Title: Incremental Web Search: Tracking Changes in the Web
Candidate: Wang, Ziyang
Advisor(s): Davis, Ernest
Abstract:
A large amount of new information is posted on the Web every day. Large-scale web search engines often update their index slowly and are unable to present such information in a timely manner. Here we present our solutions of searching new information from the web by tracking the changes of web documents.
First, we present the algorithms and techniques useful for solving the following problems: detecting web pages that have changed, extracting changes from different versions of a web page, and evaluating the significance of web changes. We propose a two-level change detector: MetaDetector and ContentDetector. The combined detector successfully reduces network traffic by about 67%. Our algorithm for extracting web changes consists of three steps: document tree construction, document tree encoding and tree matching. It has linear time complexity and extracts effectively the changed content from different versions of a web page. In order to evaluate web changes, we propose a unified ranking framework combining three metrics: popularity ranking, content-based ranking and evolution ranking. Our methods can identify and deliver important new information in a timely manner.
Second, we present an application using the techniques and algorithms we developed, named "Web Daily News Assistant (WebDNA): finding what's new on Your Web". It is a search tool that helps community users search new information on their community web. Currently WebDNA is deployed on the New York University web site.
Third, we model the changes of web documents using survival analysis. Modeling web changes is useful for web crawler scheduling and web caching. Currently people model changes to web pages as a Poisson Process, and use a necessarily incomplete detection history to estimate the true frequencies of changes. However, other features that can be used to predict change frequency have not previously been studied. Our analysis shows that PageRank value is a good predictor. Statistically, the change frequency is a function proportional to $\exp[0.36\cdot (\ln(PageRank)+C)]$. We further study the problem of combining the predictor and change history into a unified framework. An improved estimator of change frequency is presented, which successfully reduces the error by 27.3% when the change history is short.
Title: Fast Algorithms for Burst Detection
Candidate: Zhang, Xin
Advisor(s): Shasha, Dennis
Abstract:
Events occur in every aspect of our lives.
An unexpectedly large number of events occurring within some certain measurement (e.g. within some time duration or a spatial region) is called a {\em burst}, suggesting unusual behaviors or activities. Bursts come up in many natural and social processes. It is a challenging task to monitor the occurrence of bursts whose lasting duration is unknown in a fast data stream environment.
This work describes efficient data structures and algorithms for high performance burst detection under different settings. Our view is that bursts, as an unusual phenomenon, constitute a useful preliminary primitive in a knowledge discovery hierarchy. Our intent is to build a high performance primitive detection algorithm to support high-level data mining tasks.
The work starts with an algorithmic framework including a family of data structures and a heuristic optimization algorithm to choose an efficient data structure given the inputs. The advantage of this framework is that it's adaptive to different inputs. Experiments on both synthetic data and real world data show the new framework significantly outperforms existing techniques over a variety of inputs.
Furthermore, we present a greedy dynamic detection algorithm which handles the changing data. It evolves the structure to adapt to the incoming data. It achieves better performance in both synthetic and real data streams than a static algorithm in most cases.
We have applied this framework to different real world applications in physics, stock trading and website traffic monitoring. All the case studies show our framework has superb performance.
We extend this framework to multi-dimensional data and use it in an epidemiology simulation to detect infectious disease outbreak and spread.
Title: High Performance Algorithms for Multiple Streaming Time Series
Candidate: Zhao, Xiaojian
Advisor(s): Shasha, Dennis
Abstract:
Data arriving in time order (a data stream) arises in fields ranging from physics to finance to medicine to music, to name a few. Often the data comes from sensors (in physics and medicine for example) whose data rates continue to improve dramatically as sensor technology improves. Furthermore, the number of sensors is increasing, so analyzing data between sensors becomes ever more critical in order to distill knowledge from the data. Fast response is desirable in many applications (e.g. to aim a telescope at an activity of interest or to perform a stock trade). In applications such as finance, recent information, e.g. correlation, is of far more interest than older information, so analysis over sliding windows is a desired operation.
These three factors -- huge data size, fast response, and windowed computation -- motivated this work. Our intent is to build a foundational library of primitives to perform online or near online statistical analysis, e.g. windowed correlation, incremental matching pursuit, burst detection, on thousands or even millions of time series. Beside the algorithms, we also propose the concept of ``uncooperative'' time series, whose power spectra are spread over all frequencies with any regularity.
Previous work showed how to do windowed correlation with Fast Fourier Transforms and Wavelet Transforms, but such techniques don't work for uncooperative time series. This thesis will show how to use sketches (random projections) in a way that combines several simple techniques -- sketches, convolution, structured random vectors, grid structures, combinatorial design, and bootstrapping -- to achieve high performance, windowed correlation over a variety of data sets. Experiments confirm the asymptotic analysis.
To conduct matching pursuit (MP) over time series windows, an incremental scheme is designed to reduce the computational effort. Our empirical study demonstrates a substantial improvement in speed.
In previous work, Zhu and Shasha introduced an efficient algorithm to monitor bursts within windows of multiple sizes. We implemented it in a physical system by overcoming several practical challenges. Experimental results support the authors' linear running time analysis.
Title: Distribution of Route-Impacting Control Information in a Publish/Subscribe System with Delivery Guarantees
Candidate: Zhao, Yuanyuan
Advisor(s): Kedem, Zvi
Abstract:
Event-driven middleware is a popular infrastructure for building large-scale asynchronous distributed systems. Content-based publish/subscribe systems are a type of event-driven middleware that provides service flexibility and specification expressiveness, creating opportunities for improving reliability and efficiency of the system.
The use of route-impacting control information, such as subscription filters and access control rules, has the potential to enable efficient routing for applications that require selective and regional distribution of events. Such applications range from financial information systems to sensor networks to service-oriented architectures. However, it has been a great challenge to design correct and efficient protocols for distributing control information and exploiting it to achieve efficient and highly available message routing.
In this dissertation, we study the problem of distributing and utilizing route-impacting control information. We present an abstract model of content-based routing and reliable delivery in redundant broker networks. Based on this model, we design a generic algorithm that propagates control information and performs content-based routing and delivers events reliably. The algorithm is efficient and light-weight in that it does not require heavy-weight consensus protocols between redundant brokers. We extend this generic algorithm to support consolidation and merging of control information. Existing protocols can be viewed as particular encodings and optimizations of the generic algorithm. We show an encoding using virtual time vectors that supports reliable delivery and deterministic dynamic access control in redundant broker networks. In our system, the semantics of reliable delivery is clearly defined even if subscription information and access control policy can dynamically change. That is, one or more subscribers of same principal will receive exactly the same sequence of messages (modulo subscription filter differences) regardless of where they are connected and the network latency and failure conditions in their parts of the network.
We have implemented these protocols in a fully-functioning content-based publish/subscribe system - Gryphon. We evaluate its efficiency, scalability and high availability.
Title: Infrastructure for Automatic Dynamic Deployment of J2EE Applications in Distributed Environments
Author(s): Akkerman, Anatoly; Totok, Alexander; Karamcheti, Vijay
Abstract:
Recent studies showed potential for using component frameworks for building flexible adaptible applications for deployment in distributed environments. However this approach is hindered by the complexity of deployment of component-based applications, which usually involves a great deal of configuration of both the application components and system services they depend on. In this paper we propose an infrastructure for automatic dynamic deployment of J2EE applications,that specifically addresses the problems of (1) inter-component connectivity specification and its effects on component configuration and deployment; and (2) application component dependencies on application server services, their configuration and deployment. The proposed infrastructure provides simple yet expressive abstractions for potential application adaptation through dynamic deployment and undeployment of components. We implement the infrastructure as a part of the JBoss J2EE application server and test it on several sample J2EE applications.
Title: Remembrance of Experiments Past: Analyzing Time Course Datasets to Discover Complex Temporal Invariants
Author(s): Antoniotti, Marco; Ramakrishnan, Naren; Kumar, Deept; Spivak, Marina; Mishra, Bud
Abstract:
Motivation: Current microarray data analysis techniques draw the biologist's attention to targeted sets of genes but do not otherwise present global and dynamic perspectives (e.g., invariants) inferred collectively over a dataset. Such perspectives are important in order to obtain a process-level understanding of the underlying cellular machinery, especially how cells react, respond, and recover from stresses.
Results: We present GOALIE, a novel computational approach and software system that uncovers formal temporal logic models of biological processes from time course microarray datasets. GOALIE `redescribes' data into the vocabulary of biological processes and then pieces together these redescriptions into a Kripke-structure model, where possible worlds encode transcriptional states and are connected to future possible worlds. This model then supports various query, inference, and comparative assessment tasks, besides providing descriptive process-level summaries. An application of GOALIE to characterizing the yeast (S. cerevisiae) cell cycle is described.
Availability: GOALIE runs on Windows XP platforms and is available on request from the authors.
Title: An Abstract Decision Procedure for Satisfiability in the Theory of Recursive Data Types
Author(s): Barrett, Clark; Shikanian, Igor; Tinelli, Cesare
Abstract:
The theory of recursive data types is a valuable modeling tool for software verification. In the past, decision procedures have been proposed for both the full theory and its universal fragment. However, previous work has been limited in various ways, including an inability to deal with multiple constructors, multi-sorted logic, and mutually recursive data types. More significantly, previous algorithms for the universal case have been based on inefficient nondeterministic guesses and have been described in fairly complex procedural terms.
We present an algorithm which addresses these issues for the universal theory. The algorithm is presented declaratively as a set of abstract rules which are terminating, sound, and complete. We also describe strategies for applying the rules and explain why our recommended strategy is more efficient than those used by previous algorithms. Finally, we discuss how the algorithm can be used within a broader framework of cooperating decision procedures.
Title: Squidball: An Experiment in Large-Scale Motion Capture and Game Design
Author(s): Bregler, Christoph; Castiglia, Clothilde; DeVincenzo, Jessica; DuBois, Roger Luke; Feeley, Kevin; Igoe, Tom; Meyer, Jonathan; Naimark, Michael; Postelnicu, Alexandru; Rabinovich, Michael; Rosenthal, Sally; Salen, Katie; Sudol, Jeremi; Wright, Bo
Abstract:
This paper describes a new large-scale motion capture based game that is called Squidball. It was tested on up to 4000 player audiences last summer at SIGGRAPH 2004. It required to build the world's largest motion capture space, the largest motion capture markers (balls), and many other challenges in technology, production, game play, and social studies. Our aim was to entertain the SIGGRAPH Electronic Theater audience with a cooperative and energetic game that is played by everybody together, in controlling real-time graphics and audio, while bouncing and batting multiple large helium filled balloons across the entire theater space. We detail in this paper all the lessons learned in producing such a system and game, and argue why we believe Squidball was a great success.
Title: A Domain Decomposition Discretization of Parabolic Problems
Author(s): Dryja, Maksymilian; Tu, Xuemin
Abstract:
In recent years, domain decomposition methods have attracted much attention due to their successful application to many elliptic and parabolic problems. Domain decomposition methods treat problems based on a domain substructuring, which is attractive for parallel computation, due to the independence among the subdomains. In principle, domain decomposition methods may be applied to the system resulting from a standard discretization of the parabolic problems or, directly, be carried out through a direct discretization of parabolic problems. In this paper, a direct domain decomposition method is introduced to discretize the parabolic problems. The stability and convergence of this algorithm are analyzed, and an $O(\tau+h)$ error bound is provided.
Title: Translation Validation of Optimizing Compilers
Candidate: Fang, Yi
Advisor(s): Pnueli, Amir; Zuck, Lenore
Abstract:
There is a growing awareness, both in industry and academia, of the crucial role of formally verifying the translation from high-level source-code into low-level object code that is typically performed by an optimizing comiler. Formally verifying an optimizing compiler, as one woule verify any other large program, is not feasible due to its size, ongoing evolution and modification, and possibly, proprietary considerations. Translation validation is a novel approach that offers an alternative to the verification of translator in general and compilers in particular: Rather than verifying the compiler itself, one constructs a validation tool which, after every run of the compiler, formally confirms that the target code produced in the run is a correct translation of the source program. This thesis work takes an important step towards ensuring an extremely high level of confidence in compilers targeted at EPIC architectures.
In this thesis, we focus on the translation validation of structure preserving optimizations, i.e. transformations that do not modify programs' structure in a major way. This category of optimizations covers most of the global optimizations performed by compilers. This thesis has two main parts. One develops a proof rule that formally establishes the correctness of structure preserving transformation based on computational induction. The other part is the development of a tool that applies the proof rule to the automatic validation of global optimizaitons performed by Intel's ORC compiler for IA-64 architecture. With minimal instrumentation from the compiler, the tool constructs ''verification conditions'' -- formal theorems that, if valid, establish the correctness of a translation. The verificaiton conditions are then transferred to an automatic theorem prover that checks their validity. Together, the tool offers a fully automatic method to formally establish the correctness of each translation.
Title: Nonlinear Image Representation via Local Multiscale Orientation
Author(s): Hammond, David K.; Simoncelli, Eero P.
Abstract:
We present a nonlinear image representation based on multiscale local orientation measurements. Specifically, an image is first decomposed using a two-orientation steerable pyramid, a tight-frame representation in which the basis functions are directional derivatives of a radially symmetric blurring operator. The pair of subbands at each scale are thus gradients of progressively blurred copies of the original image. We then discard the magnitude information and retain only the orientation of each gradient vector. We develop a method for reconstructing the original image from this orientation information using an algorithm based on projection onto convex sets, and demonstrate its robustness to quantization.
Title: Oriented Overlays For Clustering Client Requests To Data-Centric Network Services
Author(s): He, Congchun; Karamcheti, Vijay
Abstract:
Many of the data-centric network services deployed today hold massive volumes of data at their origin websites, accessing the data to dynamically generate responses. Such dynamic responses are poorly supported by traditional caching infrastructures and result in poor performance and scalability for such services. One way of remedying this situation is to develop alternative caching infrastructures, which can dynamically detect the often large degree of service usage locality and leverage such information to on-demand replicate and redirect requests to service portions at appropriate network locations. Key to building such infrastructures is the ability to cluster and inspect client requests, at various points across a wide-area network.
This paper presents a zone-based scheme for constructing oriented overlays, which provide such an ability. Oriented overlays differ from previously proposed unstructured overlays in supporting network traffic flows from many sources towards one (or a small number) of destinations, and vice-versa. A good oriented overlay would offer sufficient clustering ability without adversely affecting path latencies. Our overlay construction scheme organizes participating nodes into different zones according to their latencies from the origin server(s), and has each node associate with one or more parents in another zone closer to the origin. Extensive experiments with a PlanetLab-based implementation of our scheme shows that it produces overlays that are (1) robust to network dynamics; (2) offer good clustering ability; and (3) minimally impact end-to-end network latencies seen by clients.
Title: An Analysis of Usage Locality for Data-Centric Web Services
Author(s): He, Congchun; Karamcheti, Vijay
Abstract:
The growing popularity of XML Web Services is resulting in a significant increase in the proportion of Internet traffic that involves requests to and responses from Web Services. Unfortunately, web service responses, because they are generated dynamically, are considered ``uncacheable" by traditional caching infrastructures. One way of remedying this situation is by developing alternative caching infrastructures, which improve performance using on-demand service replication, data offloading, and request redirection. These infrastructures benefit from two characteristics of web service traffic --- (1) the open nature of the underlying protocols, SOAP, WSDL, UDDI, which results in service requests and responses adhering to a well-formatted, widely known structure; and (2) the observation that for a large number of currently deployed data-centric services, requests can be interpreted as structured accesses against a physical or virtual database --- but require that there be sufficient locality in service usage to offset replication and redirection costs.
This paper investigates whether such locality does in fact exist in current web service workloads. We examine access logs from two large data-centric web service sites, SkyServer and TerraServer, to characterize workload locality across several dimensions: data space, network regions, and different time epochs. Our results show that both workloads exhibit a high degree of spatial and network locality: 10\% of the client IP addresses in the SkyServer trace contribute to about 99.95\% of the requests, and 99.94\% of the requests in the TerraServer trace are directed towards regions that represent less than 10\% of the overall data space accessible through the service. Our results point to the substantial opportunity for improving Web Services scalability by on-demand service replication.
Title: Translation Validation of Loop Optimizations
Candidate: Hu, Ying
Advisor(s): Goldberg, Benjamin; Barrett, Clark
Abstract:
Formal verification is important in designing reliable computer systems. For a critical software system, it is not enough to have a proof of correctness for the source code, there must also be an assurance that the compiler produces a correct translation of the source code into the target machine code. Verifying the correctness of modern optimizing compilers is a challenging task because of their size, their complexity, and their evolution over time.
In this thesis, we focus on the Translation Validation of loop optimizations. In order to validate the optimizations performed by the compiler, we try to prove the equivalence of the intermediate codes before and after the optimizations. There were previously a set of proof rules for building the equivalence relation between two programs. However, they cannot validate some cases with legal loop optimizations. We propose new proof rules to consider the conditions of loops and possible elimination of some loops, so that those cases can also be handled. According to these new proof rules, algorithms are designed to apply them to an automatic validation process.
Based on the above proof rules, we implement an automatic validation tool for loop optimizations which analyzes the loops, guesses what kinds of loop optimizations occur, proves the validity of a combination of loop optimizations, and synthesizes a series of intermediate codes. We integrate this new loop tool into our translation validation tool TVOC, so that TVOC handles not only optimizations which do not significantly change the structure of the code, but also loop optimizations which do change the structure greatly. With this new part, TVOC has succeeded in validating many examples with loop optimizations.
Speculative optimizations are the aggressive optimizations that are only correct under certain conditions that cannot be known at compile time. In this thesis, we present the theory and algorithms for validating speculative optimizations and generating the runtime tests necessary for speculative optimizations. We also provide several examples and the results of the algorithms for speculative optimizations.
Title: Two-Level Schwarz Algorithms, Using Overlapping Subregions, for Mortar Finite Element Methods
Author(s): Hyun Kim, Hyea; Widlund, Olof B.
Abstract:
Preconditioned conjugate gradient methods based on two-level overlapping Schwarz methods often perform quite well. Such a preconditioner combines a coarse space solver with local components which are defined in terms of subregions which form an overlapping covering of the region on which the elliptic problem is defined. Precise bounds on the rate of convergence of such iterative methods have previously been obtained in the case of conforming lower order and spectral finite elements as well as in a number of other cases. In this paper, this domain decomposition algorithm and analysis are extended to mortar finite elements. It is established that the condition number of the relevant iteration operator is independent of the number of subregions and varies with the relative overlap between neighboring subregions linearly as in the conforming cases previously considered.
Title: Construction of Component-Based Applications by Planning
Candidate: Kichkaylo, Tatiana
Advisor(s): Karamcheti, Vijay; Ernest Davis
Abstract:
Many modern wide-area distributed systems are component-based. This approach provides great flexibility in adapting applications to the changing state of the environment and user requirements, but increases the complexity of configuring the applications. Because of the scale and heterogeneity of modern wide-area environments, manual configuration is hard, inefficient, suboptimal, and error-prone. Automated application configuration is desired.
Constructing distributed applications requires choosing a set of components that will constitute the application instance and assigning network resources to component executions and data transfers. Stated this way, the application configuration problem (ACP) is similar to the planning (action selection) and scheduling (resource allocation) problems studied by the Artificial Intelligence (AI) community.
This thesis investigates the problem of solving the ACP using AI planning techniques. However, the ACP poses several challenges not usually encountered and addressed by the traditional AI solutions. The problem specification for the ACP can be much larger than the solution, with the relevant portions only identified during the search. Additionally, the interactions between planning operators are numeric rather than logical. Finally, it is desirable to be able to trade off quality of the solution versus search time.
We show that the ACP is undecidable in general. Therefore, instead of a single algorithm, we propose a set of techniques that can be used to compose an algorithm for a particular variety of the ACP that can exploit natural restrictions exhibited by that variety. These techniques address the challenges above by dynamically obtaining portions of the problem specification as necessary during the search, using envelope hierarchies based on numeric information for pruning and search guidance, and discretizing continuous variables to approximate numeric parameters without restricting the form of supported numeric functions.
We illustrate these techniques by describing their use in algorithms tailored for two specific varieties of the ACP --- snapshot configurations for dynamic component-based frameworks, and scheduling of grid workflows with replica selection and explicit resource reservations. Experimental evaluation of the performance of these two algorithms shows that the techniques successfully achieve their goals, with acceptable run-time overhead.
Title: A BDDC algorithm for problems with mortar discretization
Author(s): Kim, Hyea Hyun; Dryja, Maksymilian; Widlund, Olof B.
Abstract:
A BDDC (balancing domain decomposition by constraints) algorithm is developed for elliptic problems with mortar discretizations for geometrically non-conforming partitions in both two and three spatial dimensions. The coarse component of the preconditioner is defined in terms of one mortar constraint for each edge/face which is an intersection of the boundaries of a pair of subdomains. A condition number bound of the form $C \max_i \left\{ (1+\text{log} (H_i/h_i) )3 \right\}$ is established. In geometrically conforming cases, the bound can be improved to $C \max_i \left\{ (1+\text{log} (H_i/h_i) )2 \right\}$. This estimate is also valid in the geometrically nonconforming case under an additional assumption on the ratio of mesh sizes and jumps of the coefficients. This BDDC preconditioner is also shown to be closely related to the Neumann-Dirichlet preconditioner for the FETI--DP algorithms of \cite{K-04-3d,KL-02} and it is shown that the eigenvalues of the BDDC and FETI--DP methods are the same except possibly for an eigenvalue equal to 1.
Title: A FETI-DP formulation of three dimensional elasticity problems with mortar discretization
Author(s): Kim, Hyea Hyun
Abstract:
In this paper, a FETI-DP formulation for the three dimensional elasticity problem on non-matching grids over a geometrically conforming subdomain partition is considered. To resolve the nonconformity of the finite elements, a mortar matching condition on the subdomain interfaces (faces) is imposed. By introducing Lagrange multipliers for the mortar matching constraints, the resulting linear system becomes similar to that of a FETI-DP method. In order to make the FETI-DP method efficient for solving this linear system, a relatively large set of primal constraints, which include average and momentum constraints over interfaces (faces) as well as vertex constraints, is introduced. A condition number bound $C(1+\text{log}(H/h))2$ for the FETI-DP formulation with a Neumann-Dirichlet preconditioner is then proved for the elasticity problems with discontinuous material parameters when only some faces are chosen as primal faces on which the average and momentum constraints will be imposed. An algorithm which selects a quite small number of primal faces is also discussed.
Title: FETI--DP, BDDC, and Block Cholesky Methods
Author(s): Li, Jing; Widlund, Olof B.
Abstract:
Two popular non-overlapping domain decomposition methods, the FETI--DP and BDDC algorithms, are reformulated using Block Cholesky factorizations, an approach which can provide a useful framework for the design of domain decomposition algorithms for solving symmetric positive definite linear system of equations. Instead of introducing Lagrange multipliers to enforce the coarse level, primal continuity constraints in these algorithms, a change of variables is used such that each primal constraint corresponds to an explicit degree of freedom. With the new formulations of these algorithms, a simplified proof is provided that the spectra of a pair of FETI--DP and BDDC algorithms, with the same set of primal constraints, are the same. Results of numerical experiments also confirm this result.
Title: BDDC Algorithms for Incompressible Stokes Equations
Author(s): Li, Jing; Widlund, Olof B.
Abstract:
The purpose of this paper is to extend the BDDC (balancing domain decomposition by constraints) algorithm to saddle-point problems that arise when mixed finite element methods are used to approximate the system of incompressible Stokes equations. The BDDC algorithms are iterative substructuring methods, which form a class of domain decomposition methods based on the decomposition of the domain of the differential equations into nonoverlapping subdomains. They are defined in terms of a set of primal continuity constraints, which are enforced across the interface between the subdomains and which provide a coarse space component of the preconditioner. Sets of such constraints are identified for which bounds on the rate of convergence can be established that are just as strong as previously known bounds for the elliptic case. In fact, the preconditioned operator is effectively positive definite, which makes the use of a conjugate gradient method possible. A close connection is also established between the BDDC and FETI-DP algorithms for the Stokes case.
Title: On the Use of Inexact Subdomain Solvers for BDDC Algorithms
Author(s): Li, Jing; Widlund, Olof B.
Abstract:
The standard BDDC (balancing domain decomposition by constraints) preconditioner is shown to be equivalent to a preconditioner built from a partially subassembled finite element model. This results in a system of linear algebraic equations which is much easier to solve in parallel than the fully assembled model; the cost is then often dominated by that of the problems on the subdomains. An important role is also played, both in theory and practice, by an average operator and in addition exact Dirichlet solvers are used on the subdomains in order to eliminate the residual in the interior of the subdomains. The use of inexact solvers for these problems and even the replacement of the Dirichlet solvers by a trivial extension are considered. It is established that one of the resulting algorithms has the same eigenvalues as the standard BDDC algorithm, and the connection of another with the FETI-DP algorithm with a lumped preconditioner is also considered. Multigrid methods are used in the experimental work and under certain assumptions, it can be established that the iteration count essentially remains the same as when exact solvers are used, while considerable gains in the speed of the algorithm can be realized since the cost of the exact solvers grows superlinearly with the size of the subdomain problems while the multigrid methods are linear.
Title: Real-time rendering of normal maps with discontinuities
Author(s): Parilov, Evgueni; Rosenberg, Ilya; Zorin, Denis
Abstract:
Title: Real-time rendering of normal maps with discontinuities (NYU-CS-TR872) Authors: Evgueni Parilov, Ilya Rosenberg and Denis Zorin Abstract:
Normal mapping uses normal perturbations stored in a texture to give objects a more geometrically complex appearance without increasing the number of geometric primitives. Standard bi- and trilinear interpolation of normal maps works well if the normal field is continuous, but may result in visible artifacts in the areas where the field is discontinuous, which is common for surfaces with creases and dents.
In this paper we describe a real-time rendering technique which preserves the discontinuity curves of the normal field at sub-pixel level and its GPU implementation. Our representation of the piecewise-continuous normal field is based on approximations of the distance function to the discontinuity set and its gradient. Using these approximations we can efficiently reconstruct discontinuities at arbitrary resolution and ensure that no normals are interpolated across the discontinuity. We also described a method for updating the normal field along the discontinuities in real-time based on blending the original field with the one calculated from a user-defined surface profile.
Title: Algorithmic Algebraic Model Checking I: The Case of Biochemical Systems and their Reachability Analysis
Author(s): Piazza, C.; Antoniotto, M.; Mysore, V.; Policriti, A.; Winkler, F.; Mishra, B.
Abstract:
Presently, there is no clear way to determine if the current body of biological facts is sufficient to explain phenomenology. Rigorous mathematical models with automated tools for reasoning, simulation, and computation can be of enormous help to uncover cognitive flaws, qualitative simplification or overly generalized assumptions. The approaches developed by control theorists analyzing stability of a system with feedback, physicists studying asymptotic properties of dynamical systems, computer scientists reasoning about discrete or hybrid (combining discrete events with continuous events) reactive systems---all have tried to address some aspects of the same problem in a very concrete manner. We explore here how biological processes could be studied in a similar manner, and how the appropriate tools for this purpose can be created.
In this paper, we suggest a possible confluence of the theory of hybrid automata and the techniques of algorithmic algebra to create a computational basis for systems biology. We start by discussing our basis for this choice -- semi-algebraic hybrid systems, as we also recognize its power and limitations. We explore solutions to the bounded-reachability problem through symbolic computation methods, applied to the descriptions of the traces of the hybrid automaton. Because the description of the automaton is through semi-algebraic sets, the evolution of the automaton can be described even in cases where system parameters and initial conditions are unspecified. Nonetheless, semialgebraic decision procedures provide a succinct description of algebraic constraints over the initial values and parameters for which proper behavior of the system can be expected. In addition, by keeping track of conservation principles in terms of constraint or invariant manifolds on which the system must evolve, we avoid many of the obvious pitfalls of numerical approaches.
Title: Extensible MultiModal Environment Toolkit (EMMET): A Toolkit for Prototyping and Remotely Testing Speech and Gesture Based Multimodal Interfaces
Candidate: Robbins, Christopher A.
Advisor(s): Perlin, Ken
Abstract:
Ongoing improvements to the performance and accessibility of less conventional input modalities such as speech and gesture recognition now provide new dimensions for interface designers to explore. Yet there is a scarcity of commercial applications which utilize these modalities either independently or multimodally. This scarcity partially results from a lack of development tools and design guidelines to facilitate the use of speech and gesture.
An integral aspect of the user interface design process is the ability to easily evaluate various design solutions through an iterative process of prototyping and testing. Through this process guidelines emerge that aid in the design of future interfaces. Today there is no shortage of tools supporting the development of conventional interfaces. However there do not exist resources allowing interface designers to easily prototype and quickly test, via remote distribution, interface designs utilizing speech and gesture.
The thesis work for this dissertation explores the development of an Extensible MultiModal Environment Toolkit (EMMET) for prototyping and remotely testing speech and gesture based multimodal interfaces to three-dimensional environments. The overarching goals for this toolkit are to allow its users to: explore speech and gesture based interface design without requiring an understanding of the details involved in the low-level implementation of speech or gesture recognition, quickly distribute their multimodal interface prototypes via the Web, and receive multimodal usage statistics collected remotely after each use of their application.
EMMET ultimately contributes to the field of multimodal user interface design by providing an environment to existing user interface developers in which speech and gesture recognition have been seamlessly integrated into their palette of user input options. Such seamless integration serves to increase the utilization within applications of speech and gesture modalities by removing any actual or perceived deterrents to the use of these modalities versus the use of conventional modalities. EMMET additionally strives to improve the quality of speech and gesture based interfaces by supporting the prototype-and-test development cycle through its Web distribution and usage statistics collection capabilities. These capabilities also allow developers to realize new design guidelines specific to the use of speech and gesture.
Title: Ranking with a P-norm Push
Author(s): Rudin, Cynthia
Abstract:
We are interested in supervised ranking with the following twist: our goal is to design algorithms that perform especially well near the top of the ranked list, and are only required to perform sufficiently well on the rest of the list. Towards this goal, we provide a general form of convex objective that gives high-scoring examples more importance. This ``push'' near the top of the list can be chosen arbitrarily large or small. We choose $\ell_p$-norms to provide a specific type of push; as $p$ becomes large, the algorithm concentrates harder near the top of the list.
We derive a generalization bound based on the $p$-norm objective. We then derive a corresponding boosting-style algorithm, and illustrate the usefulness of the algorithm through experiments on UCI data.
Title: Better Burst Detection
Author(s): Shasha, Dennis; Zhang, Xin
Abstract:
A burst is a large number of events occurring within a certain time window. As an unusual activity, it's a noteworthy phenomenon in many natural and social processes. Many data stream applications require the detection of bursts across a variety of window sizes. For example, stock traders may be interested in bursts having to do with institutional purchases or sales that are spread out over minutes or hours. Detecting a burst over any of $k$ window sizes, a problem we call {\em elastic burst detection}, in a stream of length $N$ naively requires $O(kN)$ time. Previous work \cite{DiscoveryBook03} showed that a simple Shifted Binary Tree structure can reduce this time substantially (in very favorable cases near to $O(N)$) by filtering away obvious non-bursts. Unfortunately, for certain data distributions, the filter marks many windows of events as possible bursts, even though a detailed check shows them to be non-bursts.
In this paper, we present a new algorithmic framework for elastic burst detection: a family of data structures that generalizes the Shifted Binary Tree. We then present a heuristic search algorithm to find an efficient structure among the many offered by the framework, given the input. We study how different inputs affect the desired structures. Experiments on both synthetic and real world data show a factor of up to 35 times improvement compared with the Shifted Binary Tree over a wide variety of inputs, depending on the data distribution. We show an example application that identifies interesting correlations between bursts of activity in different stocks.
Title: Modeling Of Concurrent Web Sessions With Bounded Inconsistency In Shared Data
Author(s): Totok, Alexander; Karamcheti, Vijay
Abstract:
Client interactions with modern web-accessible network services are typically organized into sessions involving multiple requests that read and write shared application data. Therefore when executed concurrently, web sessions may invalidate each other's data. Depending on the nature of the business represented by the service, allowing the session with invalid data to progress might lead to financial penalties for the service provider, while blocking the session's progress and deferring its execution (e.g., by relaying its handling to the customer service) will most probably result in user dissatisfaction. A compromise would be to tolerate some bounded data inconsistency, which would allow most of the sessions to progress, while limiting the potential financial loss incurred by the service. In order to quantitatively reason about these tradeoffs, the service provider can benefit from models that predict metrics, such as the percentage of successfully completed sessions, for a certain degree of tolerable data inconsistency.
This paper develops such analytical models of concurrent web sessions with bounded inconsistency in shared data for three popular concurrency control algorithms. We illustrate our models using the sample buyer scenario from the TPC-W e-Commerce benchmark, and validate them by showing their close correspondence to measured results of concurrent session execution in both a simulated and a real web server environment. Our models take as input parameters of service usage, which can be obtained through profiling of incoming client requests. We augment our web application server environment with a profiling and automated decision making infrastructure which is shown to successfully choose, based on the specified performance metric, the best concurrency control algorithm in real time in response to changing service usage patterns.
Title: Pattern Discovery for Hypotheses Generation in Biology
Candidate: Tsirigos, Aristotelis
Advisor(s): Shasha, Dennis
Abstract:
In recent years, the increase in the amounts of available genomic as well as gene expression data has provided researchers with the necessary information to train and test various models of gene origin, evolution, function and regulation. In this thesis, we present novel solutions to key problems in computational biology that deal with nucleotide sequences (horizontal gene transfer detection), amino-acid sequences (protein sub-cellular localization prediction), and gene expression data (transcription factor - binding site pair discovery). Different pattern discovery techniques are utilized, such as maximal sequence motif discovery and maximal itemset discovery, and combined with support vector machines in order to achieve significant improvements against previously proposed methods.
Title: A BDDC Algorithm for Mixed Formulation of Flow in Porous Media
Author(s): Tu, Xuemin
Abstract:
The BDDC (balancing domain decomposition by constraints) algorithms are similar to the balancing Neumann-Neumann methods, with a small number of continuity constraints enforced across the interface throughout the iterations. These constraints form a coarse, global component of the preconditioner. The BDDC methods are powerful for solving large sparse linear algebraic systems arising from discretizations of elliptic boundary value problems. In this paper, the BDDC algorithm is extended to saddle point problems generated from the mixed finite element methods used to approximate the scalar elliptic problems for flow in porous media.
Edge/face average constraints are enforced and the same rate of convergence is obtained as for simple elliptic cases. The condition number bound is estimated and numerical experiments are discussed. In addition, a comparison of the BDDC method with an edge/face-based iterative substructuring method is provided.
Title: BDDC Domain Decomposition Algorithms: Methods with Three Levels and for Flow in Porous Media
Author(s): Tu, Xuemin
Abstract:
Two inexact coarse solvers for Balancing Domain Decomposition by Constraints (BDDC) algorithms are introduced and analyzed. These solvers help remove a bottleneck for the two-level BDDC algorithms related to the cost of the coarse problem when the number of subdomains is large. At the same time, a good convergence rate is maintained.
BDDC algorithms are also developed for the linear systems arising from flow in porous media discretized with mixed and hybrid finite elements. Our methods are proven to be scalable and the condition numbers of the operators with our BDDC preconditioners grow only polylogarithmically with the size of the subdomain problems.
Title: A BDDC algorithm for flow in porous media with a hybrid finite element discretization
Author(s): Tu, Xuemin
Abstract:
The BDDC (balancing domain decomposition by constraints) methods have been applied successfully to solve the large sparse linear algebraic systems arising from conforming finite element discretizations of elliptic boundary value problems. In this paper, the scalar elliptic problems for flow in porous media are discretized by a hybrid finite element method which is equivalent to a nonconforming finite element method. The BDDC algorithm is extended to these problems which originate as saddle point problems.
Edge/face average constraints are enforced across the interface and the same rate of convergence is obtained as in conforming cases. The condition number of the preconditioned system is estimated and numerical experiments are discussed.
Title: Three-Level BDDC in Three Dimensions
Author(s): Tu, Xuemin
Abstract:
BDDC methods are nonoverlapping iterative substructuring domain decomposition methods for the solution of large sparse linear algebraic systems arising from discretization of elliptic boundary value problems. Its coarse problem is given by a small number of continuity constraints which are enforced across the interface. The coarse problem matrix is generated and factored by direct solvers at the beginning of the computation and it can ultimately become a bottleneck, if the number of subdomains is very large.
In this paper, two three-level BDDC methods are introduced for solving the coarse problem approximately in three dimensions. This is an extension of previous work for the two dimensional case and since vertex constraints alone do not suffice to obtain polylogarithmic condition number bound, edge constraints are considered in this paper. Some new technical tools are then needed in the analysis and this makes the three dimensional case more complicated than the two dimensional case.
Estimates of the condition numbers are provided for two three-level BDDC methods and numerical experiments are also discussed.
Title: Automatic Verification of Parameterized Systems
Candidate: Xu, Jiazhao
Advisor(s): Pnueli, Amir
Abstract:
Verification plays an indispensable role in designing reliable computer hardware and software systems. With the fast growth in design complexity and the quick turnaround in design time, formal verification has become an increasingly important technology for establishing correctness as well as for finding difficult bugs. Since there is no ``silver-bullet'' to solve all verification problems, a spectrum of powerful techniques in formal verification have been developed to tackle different verification problems and complexity issues. Depending on the nature of the problem whose most salient components are the system implementation and the property specification, a proper methodology or a combination of different techniques is applied to solve the problem.
In this thesis, we focus on the research and development of formal methods to uniformly verify parameterized systems. A parameterized system is a class of systems obtained by instantiating the system parameters. Parameterized verification seeks a single correctness proof of a property for the entire class. Although the general parameterized verification problem is undecidable [AK86], it is possible to solve special classes by applying a repertoire of techniques and heuristics. Many methods in parameterized verification require a great deal of human interaction. This makes the application of these methods to real world problems infeasible. Thus, the main focus of this research is to develop techniques that can be automated to deliver proofs of safety and liveness properties.
Our research combines various formal techniques such as deductive methods, abstraction and model checking. One main result in this thesis is an automatic deductive method for parameterized verification. We apply small model properties of Bounded Data Systems (a special type of parameterized system) to help prove deductive inference rules for the safety properties of BDS systems. Another methodology we developed enables us to prove liveness properties of parameterized systems via an automatic abstraction method called counter abstraction . There are several useful by-products from our research: A set of heuristics is established for the automatic generation of program invariants which can benefit deductive verification in general; also we proposed methodologies for the automatic abstraction of fairness conditions that are crucial for proving liveness properties.
Title: Mobility, Route Caching, and TCP Performance in Mobile Ad Hoc Networks
Candidate: Yu, Xin
Advisor(s): Johnson, David B.
Abstract:
In a mobile ad hoc network, mobile nodes communicate with each other through wireless links. Mobility causes frequent topology changes. This thesis addresses the fundamental challenges mobility presents to on-demand routing protocols and to TCP.
On-demand routing protocols use route caches to make routing decisions. Due to mobility, cached routes easily become stale. To address the cache staleness issue, prior work used adaptive timeout mechanisms. However, heuristics cannot accurately estimate timeouts because topology changes are unpredictable. I propose to proactively disseminate the broken link information to the nodes that have cached the link. I define a new cache structure called a cache table to maintain the information necessary for cache updates, and design a distributed cache update algorithm. This algorithm is the first work that proactively updates route caches in an adaptive manner. Simulation results show that proactive cache updating is more efficient than adaptive timeout mechanisms. I conclude that proactive cache updating is key to the adaptation of on-demand routing protocols to mobility.
TCP does not perform well in mobile ad hoc networks. Prior work provided link failure feedback to TCP so that it can avoid invoking congestion control mechanisms for packet losses caused by route failures. Simulation results show that my cache update algorithm significantly improves TCP throughput since it reduces the effect of mobility on TCP. TCP still suffers from frequent data and ACK losses. I propose to make routing protocols aware of lost TCP packets and help reduce TCP timeouts. I design two mechanisms that exploit cross-layer information awareness: early packet loss notification (EPLN) and best-effort ACK delivery (BEAD). EPLN notifies TCP senders about lost data. BEAD retransmits ACKs at intermediate nodes or at TCP receivers. Simulation results show that the two mechanisms significantly improve TCP throughput. I conclude that cross-layer information awareness is key to making TCP efficient in the presence of mobility.
I also study the impact of route caching strategies on the scalability of on-demand routing protocols with mobility. I show that making route caches adapt quickly and efficiently to topology changes is key to the scalability of on-demand routing protocols with mobility.
Title: Information Extraction from Multiple Syntactic Sources
Candidate: Zhao, Shubin
Advisor(s): Grishman, Ralph
Abstract:
Information Extraction is the automatic extraction of facts from text, which includes detection of named entities, entity relations and events. Conventional approaches to Information Extraction try to find syntactic patterns based on deep processing of text, such as partial or full parsing. The problem these solutions have to face is that as deeper analysis is used, the accuracy of the result decreases, and one cannot recover from the induced errors. On the other hand, lower level processing is more accurate and it can also provide useful information. However, within the framework of conventional approaches, this kind of information can not be efficiently incorporated.
This thesis describes a novel supervised approach based on kernel methods to address these issues. In this approach customized kernels are used to match syntactic structures produced from different preprocessing phases. Using properties of a kernel, individual kernels are combined into composite kernels to integrate and extend all the information. The composite kernels can be used with various classifiers, such as Nearest Neighbor or Support Vector Machines (SVM). The main classifier we propose to use is SVM due to its ability to generalize in large dimensional feature spaces. We will show that each level of syntactic information can contribute to IE tasks, and low level information can help to recover from errors in deep processing.
The new approach has demonstrated state-of-the-art performance on two benchmark tasks. The first task is detecting slot fillers for management succession events (MUC-6). For this task two types of kernels were designed, a surface kernel based on word n-grams and a kernel built on sentence dependency trees; the second task is the ACE RDR evaluation, which is to recognize relations between entities in text from newswire and broadcast news transcript. For this task, five kernels were built to represent information from sentence tokenization, syntactic parsing and dependency parsing. Experimental results for the two tasks will be shown and discussed.
Title: Fast and Cheap Genome wide Haplotype Construction via Optical Mapping
Author(s): Anantharaman, Thomas; Mysore, Venkatesh; Mishra, Bud
Abstract:
We describe an efficient algorithm to construct genome wide haplotype restriction maps of an individual by aligning single molecule DNA fragments collected with Optical Mapping technology. Using this algorithm and small amount of genomic material, we can construct the parental haplotypes for each diploid chromosome for any individual, one from the father and the other from the mother. Since such haplotype maps reveal the polymorphisms due to single nucleotide differences (SNPs) and small insertions and deletions (RFLPs), they are useful in association studies, studies involving genomic instabilities in cancer, and genetics. For instance, such haplotype restriction maps of individuals in a population can be used in association studies to locate genes responsible for genetics diseases with relatively low cost and high throughput. If the underlying problem is formulated as a combinatorial optimization problem, it can be shown to be NP-complete (a special case of K-population problem). But by effectively exploiting the structure of the underlying error processes and using a novel analog of the Baum-Welch algorithm for HMM models, we devise a probabilistic algorithm with a time complexity that is linear in the number of markers. The algorithms were tested by constructing the first genome wide haplotype restriction map of the microbe T. Pseudoana, as well as constructing a haplotype restriction map of a 120 Megabase region of Human chromosome 4. The frequency of false positives and false negatives was estimated using simulated data. The empirical results were found very promising.
Title: Naturally Speaking: A Systems Biology Tool with Natural Language Interfaces
Author(s): Antoniotti, Marco; Lau, Ian T.; Mishra, Bud
Abstract:
This short paper describes a systems biology software tool that can engage in a dialogue with a biologist by responding to questions posed to it in English (or another natural language) regarding the behavior of a complex biological system, and by suggesting a set of facts about the biological system based on a timetested generate and test approach. Thus, this bioinformatics system improves the quality of the interaction that a biologist can have with a system built on rigorous mathematical modeling, but without being aware of the underlying mathematically sophisticated concepts or notations. Given the nature of the mathematical semantics of our Simpathica/XSSYS tool, it was possible to construct a well-founded natural language interface on top of the computational kernel. We discuss our tool and illustrate its use with a few examples. The natural language subsystem is available as an integrated subsystem of the Simpathica/XSSYS tool and through a simple Web-based interface; we describe both systems in the paper. More details about the system can be found at: http://bioinformatics.nyu.edu, and its sub-pages.
Title: Practical Packrat Parsing
Author(s): Grimm, Robert
Abstract:
A considerable number of research projects are exploring how to extend object-oriented programming languages such as Java with, for example, support for generics, multiple dispatch, or pattern matching. To keep up with these changes, language implementors need appropriate tools. In this context, easily extensible parser generators are especially important because parsing program sources is a necessary first step for any language processor, be it a compiler, syntax-highlighting editor, or API documentation generator. Unfortunately, context-free grammars and the corresponding LR or LL parsers, while well understood and widely used, are also unnecessarily hard to extend. To address this lack of appropriate tools, we introduce Rats!, a parser generator for Java that supports easily modifiable grammars and avoids the complexities associated with altering LR or LL grammars. Our work builds on recent research on packrat parsers, which are recursive descent parsers that perform backtracking but also memoize all intermediate results (hence their name), thus ensuring linear-time performance. Our work makes this parsing technique, which has been developed in the context of functional programming languages, practical for object-oriented languages. Furthermore, our parser generator supports simpler grammar specifications and more convenient error reporting, while also producing better performing parsers through aggressive optimizations. In this paper, we motivate the need for more easily extensible parsers, describe our parser generator and its optimizations in detail, and present the results of our experimental evaluation.
Title: Partitionable Services Framework: Seamless Access to Distributed Applications
Candidate: Ivan, Anca
Advisor(s): Karamcheti, Vijay
Abstract:
A key problem in contemporary distributed systems is how to satisfy user quality of service (QoS) requirements for distributed applications deployed in heterogeneous, dynamically changing environments spanning multiple administrative domains.
An attractive solution is to create an infrastructure which satisfies user QoS requirements by automatically and transparently adapting distributed applications to any environment changes with minimum user input. However, successful use of this approach requires overcoming three challenges: (1) Capturing the application behavior and its relationship with the environment as a set of compact local specifications, using both general, quantitative (e.g., CPU usage) and qualitative (e.g., security) properties. Such information should be sufficient to reason about the global behavior of the application deployment. (2) Finding the ``best'' application deployment that satisfies both application and user requirements, and the various domain policies. The search algorithm should be complete, efficient, scalable with regard to application and network sizes, and guarantee optimality (e.g., resources consumed by applications). (3) Ensuring that the found deployments are practical and efficient, i.e., that the efficiency of automatic deployments is comparable with the efficiency of hand-tuned solutions.
This dissertation describes three techniques that address these challenges in the context of component-based applications. The modularity and reusability of the latter enable automatic deployments while supporting reasoning about the global connectivity based on the local information exposed by each component. The first technique extends the basic component-based application model with information about conditions and effects of component deployments and linkages, together with interactions between components and the network. The second technique uses AI planning to build an efficient and scalable algorithm which exploits the expressivity of the application model to find an application deployment that satisfies user QoS and application requirements. The last technique ensures that application deployments are both practical and efficient, by leveraging language and run-time system support to automatically customize components, as appropriate for the desired security and data consistency guarantees. These techniques are implemented as integral parts of the Partitionable Services Framework (PSF), a Java-based framework which flexibly assembles component-based applications to suit the properties of their environment. PSF facilitates on-demand, transparent migration and replication of application components at locations closer to clients, while retaining the illusion of a monolithic application.
The benefits of PSF are evaluated by deploying representative component-based applications in an environment simulating fast and secure domains connected by slow and insecure links. Analysis of the programming and the deployment processes shows that: (1) the code modifications required by PSF are minimal,(2) PSF appropriately adapts the deployments based on the network state and user QoS requirements, (3) the run-time deployment overheads incurred by PSF are negligible compared to the application lifetime, and (4) the efficiency of PSF-deployed applications matches that of hand-crafted solutions.
Title: Sekitei: An AI planner for Constrained Component Deployment in Wide-Area Networks
Author(s): Kichkaylo, Tatiana; Ivan, Anca; Karamcheti, Vijay
Abstract:
Wide-area network applications are increasingly being built using component-based models, enabling integration of diverse functionality in modules distributed across the network. In such models, dynamic component selection and deployment enables an application to flexibly adapt to changing client and network characteristics, achieve loadbalancing, and satisfy QoS requirements. Unfortunately, the problem of finding a valid component deployment is hard because one needs to decide on the set of components while satisfying various constraints resulting from application semantic requirements, network resource limitations, and interactions between the two. In this paper, we describe a general model for the component placement problem and present an algorithm for it, which is based on AI planning algorithms. We validate the effectiveness of our algorithm by demonstrating its scalability with respect to network size and number of components in the context of deployments generated for two example applications a security-sensitive mail service, and a webcast service in a variety of network environments.
Title: Dual-Primal FETI Methods for Linear Elasticity
Author(s): Klawonn, Axel; Widlund, Olof B.
Abstract:
Dual-Primal FETI methods are nonoverlapping domain decomposition methods where some of the continuity constraints across subdomain boundaries are required to hold throughout the iterations, as in primal iterative substructuring methods, while most of the constraints are enforced by Lagrange multipliers, as in one-level FETI methods. The purpose of this article is to develop strategies for selecting these constraints, which are enforced throughout the iterations, such that good convergence bounds are obtained, which are independent of even large changes in the stiffnesses of the subdomains across the interface between them. A theoretical analysis is provided and condition number bounds are established which are uniform with respect to arbitrarily large jumps in the Young's modulus of the material and otherwise only depend polylogarithmically on the number of unknowns of a single subdomain.
Title: VALIS: A Multi-language System for Rapid Prototyping in Computational Biology
Candidate: Paxia, Salvatore
Advisor(s): Mishra, Bud
Abstract:
Bioinformatics is a challenging area for computer science, since the underlying computational formalisms span database systems, numerical methods, geometric modeling and visualization, imaging and image analysis, combinatorial algorithms, data analysis and mining, statistical approaches, and reasoning under uncertainty.
This thesis describes the Valis environment for rapid application prototyping in bioinformatics. The core components of the Valis system are the underlying database structure and the algorithmic development platform.
This thesis presents a novel set of data structures that has marked advantages when dealing with unstructured and unbounded data that are common in scientific fields and bioinformatics.
Bioinformatics problems rarely have a one-language, one-platform solution. The Valis environment allows seamless integration between scripts written in different programming languages and includes tools to rapidly prototype graphical user interfaces.
To date the speed of computation of most whole genome analysis tools have stood in the way of developing fast interactive programs that may be used as exploratory tools. This thesis presents the basic algorithms and widgets that permit rapid prototyping of whole genomic scale real-time applications within Valis.
Title: Thick Surfaces: Interactive Modeling of Topologically Complex Geometric Details
Candidate: Peng, Jianbo
Advisor(s): Zorin, Denis
Abstract:
Lots of objects in computer graphics applications are represented by surfaces. It works very well for objects of simple topology, but can get prohibitively expensive for objects with complex small-scale geometrical details.
Volumetric textures aligned with a surface can be used to add topologically complex geometric details to an object, while retaining an underlying simple surface structure. The simple surface structure provides great controllability on the overall shape of the model, and volumetric textures handle geometric details and topological changes efficiently.
Adding a volumetric texture to a surface requires more than a conventional twodimensional parameterization: a part of the space surrounding the surface has to be parameterized. Another problem with using volumetric textures for adding geometric detail is the difficulty of the rendering of implicitly represented surfaces, especially when they are changed interactively.
We introduce thick surfaces to represent objects with topologically complex geometric details. A thick surface consists of three components. First, a base mesh of simple structure is used to approximate the overall shape of the object. Second, a layer of space along the base mesh is parameterized. We define the layer of space as a shell, which covers the geometric details of the object. Third, volumetric textures of geometric details are mapped into the shell. The object is represented as the implicit surface encoded by the volumetric textures. Places without volumetric textures are filled with patches of the base mesh.
We present algorithms for constructing a shell around a surface and rendering a volumetric-textured surface. Mipmap technique for volumetric textures is explored as well. The gradient field of a generalized distance function is used to construct a non-self-intersecting shell, which has other properties desirable for volumetric texture mapping. The rendering algorithm is designed and implemented on NVIDIA GeForceFX video chips. Finally we demonstrate a number of interactive operations that these algorithms enable.
Title: TM-LPSAT: Encoding Temporal Metric Planning in Continuous Time
Candidate: Shin, Ji-Ae
Advisor(s): Davis, Ernest
Abstract:
In any domain with change, the dimension of time is inherently involved. Whether the domain should be modeled in discrete time or continuous time depends on aspects of the domain to be modeled. Many complex real-world domains involve continuous time, resources, metric quantities and concurrent actions. Planning in such domains must necessarily go beyond simple discrete models of time and change.
In this thesis, we show how the SAT-based planning framework can be extended to generate plans of concurrent asynchronous actions that may depend on or make change piecewise linear metric constraints in continuous time.
In the SAT-based planning framework, a planning problem is formulated as a satisfiability problem of a set of propositional constraints (axioms) such that any model of the axioms corresponds to a valid plan. There are two parameters to a SAT-based planning system: an encoding scheme for representing plans of bounded length and a propositional SAT solver to search for a model. The LPSAT architecture is composed of a SAT solver integrated with a linear arithmetic constraint solver in order to deal with metric aspects of domains.
We present encoding schemes for temporal models of continuous time defined in PDDL+: ( i ) Durative actions with discrete and/or continuous changes; (ii) Real-time temporal model with exogenous events and autonomous processes capturing continuous changes. The encoding represents, in a CNF formula over arithmetic constraints and propositional fluents, time-stamped parallel plans possibly with concurrent continuous and/or discrete changes. In addition, we present encoding schemes for multi-capacity resources, partitioned interval resources, and metric quantities which are represented as intervals. An interval type can be used as a parameter to action as well as a fluent type.
Based on the LPSAT engine, the TM-LPSAT temporal metric planner has been implemented: Given a PDDL+ representation of a planning problem, the compiler of TM-LPSAT translates it in a CNF formula, which is fed into the LPSAT engine to find a solution corresponding to a plan for the planning problem. We also have experimented on our temporal metric encodings with other decision procedure, MathSAT, which deals with propositional combinations of linear constraints and Boolean variables. The results show that in terms of searching time the SAT-based approach to temporal metric planning can be comparable to other planning approaches and there is plenty of room to push further the limits of the SAT-based approach.
Title: Optical flow estimation as distributed optimization problem - an aVLSI implementation
Author(s): Stocker, Alan
Abstract:
I present a new focal-plane analog VLSI sensor that estimates optical flow in two visual dimensions. The chip significantly improves previous approaches both with respect to the applied model of optical flow estimation as well as the actual hardware implementation. Its distributed computational architecture consists of an array of locally connected motion units that collectively solve for the unique optimal optical flow estimate. The novel gradient-based motion model assumes visual motion to be translational, smooth and biased. The model guarantees that the estimation problem is computationally well-posed regardless of the visual input. Model parameters can be globally adjusted, leading to a rich output behavior. Varying the smoothness strength, for example, can provide a continuous spectrum of motion estimates, ranging from normal to global optical flow. Unlike approaches that rely on the explicit matching of brightness edges in space or time, the applied gradient-based model assures spatiotemporal continuity on visual information. The non-linear coupling of the individual motion units improves the resulting optical flow estimate because it reduces spatial smoothing across large velocity differences. Extended measures of a 30x30 array prototype sensor under real-world conditions demonstrate the validity of the model and the robustness and functionality of the implementation.
Title: Unsupervised Discovery of Extraction Patterns for InformationExtraction
Candidate: Sudo, Kiyoshi
Advisor(s): Grishman, Ralph; Sekine, Satoshi
Abstract:
The task of Information Extraction (IE) is to find specific types of information in natural language text. In particular, *event extraction* identifies instances of a particular type of event or fact (a particular "scenario"), including the entities involved, and fills a database which has been pre-defined for the scenario. As the number of documents available on-line has multiplied, entity extraction has grown in importance for various applications, including tracking terrorist activities from newswire sources and building a database of job postings from the Web, to name a few.
Linguistic contexts, such as predicate-argument relationships, have been widely used as *extraction patterns* to identify the items to be extracted from the text. The cost of creating extraction patterns for each scenario has been a bottleneck limiting the portability of information extraction systems to different scenarios, although there has been some research on semi-supervised pattern discovery procedures to reduce this cost. The challenge is to develop a fully automatic method for identifying extraction patterns for a scenario specified by the user.
This dissertation presents a novel approach for the unsupervised discovery of extraction patterns for event extraction from raw text. First, we present a framework that allows the user to have a self-customizing information extraction system for his/her query: the Query-Driven Information Extraction (QDIE) framework. The input to the QDIE framework is the user's query: either a set of keywords or a narrative description of the event extraction task.
Second, we assess the improvement in extraction pattern models. By considering the shortcomings of the prior work based on predicate-argument models and their extensions, we propose a novel extraction pattern model that is based on arbitrary subtrees of dependency trees.
Third, we address the issue of portability across languages. As a case study of the QDIE framework, we implemented a pre-CODIE system, a Cross-Lingual On-Demand Information Extraction system requiring minimal human intervention, which incorporates the QDIE framework as a component for pattern discovery. In addition, we assess the role of machine translation in cross-lingual information extraction by comparing translation-based implementations.
Title: Three-level BDDC in Two Dimensions
Author(s): Tu, Xuemin
Abstract:
BDDC methods are nonoverlapping iterative substructuring domain decomposition methods for the solutions of large sparse linear algebraic systems arising from discretization of elliptic boundary value problems. They are similar to the balancing Neumann-Neumann algorithm. However, in BDDC methods, a small number of continuity constraints are enforced across the interface, and these constraints form a new coarse, global component. An important advantage of using such constraints is that the Schur complements that arise in the computation willa ll be strictly positive definite. The coarse problem is generated and factored by a direct solver at the beginning of the computation. However, this problem can ultimately become a bottleneck, if the number of subdomains is very large. In this paper, two three-level BDDC methods are introduced for solving the coarse problem approximately in two dimensional space, while still maintaining a good convergence rate. Estimates of the condition numbers are provided for the two three-level BDDC methods and numerical experiments are also discussed.
Title: An Efficient and High-Order Accurate Boundary Integral Solver for the Stokes Equations in Three Dimensional Complex Geometries
Candidate: Ying, Lexing
Advisor(s): Zorin, Denis
Abstract:
This dissertation presents an efficient and high-order boundary integral solver for the Stokes equations in complex 3D geometries. The targeted applications of this solver are the flow problems in domains involving moving boundaries. In such problems, traditional finite element methods involving 3D unstructured mesh generation expe- rience difficulties. Our solver uses the indirect boundary integral formulation and discretizes the equation using the Nyström method.
Although our solver is designed for the Stokes equations, we show that it can be generalized to other constant coefficient elliptic partial differential equations (PDEs) with non-oscillatory kernels.
First, we present a new geometric representation of the domain boundary. This scheme takes quadrilateral control meshes with arbitrary geometry and topology as input, and produces smooth surfaces approximating the control meshes. Our surfaces are parameterized over several overlapping charts through explicit nonsingular C ^{ ∞ } parameterizations, depend linearly on the control points, have fixed-size local support for basis functions, and have good visual quality.
Second, we describe a kernel independent fast multipole method (FMM) and its parallel implementation. The main feature of our algorithm is that it is based only on kernel evaluation and does not require the multipole expansions of the underlying kernel. We have tested our method on kernels from a wide range of elliptic PDEs. Our numerical results indicate that our method is efficient and accurate. Other ad- vantages include the simplicity of the implementation and its immediate extension to other elliptic PDE kernels. We also present an MPI based parallel implementation which scales well up to thousands of processors.
Third, we present a framework to evaluate the singular integrals in our solver. A singular integral is decomposed into a smooth far field part and a local part that contains the singularity. The smooth part of the integral is integrated using the trape- zoidal rule over overlapping charts, and the singular part is integrated in the polar coordinates which removes or decreases the order of singularity. We also describe a novel algorithm to integrate the nearly singular integrals coming from the evaluation at points close to the boundary.
Title: High Performance Data Mining in Time Series: Techniques and Case Studies
Candidate: Zhu, Yunyue
Advisor(s): Shasha, Dennis
Abstract:
Note: A significantly improved and expanded description of this material is available in the book High Performance Discovery in Time Series Springer Verlag 2004 ISBN 0-387-00857-8.
As extremely large time series data sets grow more prevalent in a wide variety of settings, we face the significant challenge of developing efficient analysis methods. This dissertation addresses the problem in designing fast, scalable algorithms for the analysis of time series.
The first part of this dissertation describes the framework for high performance time series data mining based on important primitives. Data reduction trasform such as the Discrete Fourier Transform, the Discrete Wavelet Transform, Singular Value Decomposition and Random Projection, can reduce the size of the data without substantial loss of information, therefore provides a synopsis of the data. Indexing methods organize data so that the time series data can be retrieved efficiently. Transformation on time series, such as shifting, scaling, time shifting, time scaling and dynamic time warping, facilitates the discovery of flexible patterns from time series.
The second part of this dissertation integrates the above primitives into useful applications ranging from music to physics to finance to medicine.
StatStream
StatStream is a system based on fast algorithms for finding the most highly correlated pairs of time series from among thousands of time series streams and doing so in a moving window fashion. It can be used to find correlations in time series in finance and in scientific applications.
HumFinder
Most people hum rather poorly. Nevertheless, somehow people have some idea what we are humming when we hum. The goal of the query by humming program, HumFinder, is to make a computer do what a person can do. Using pitch translation, time dilation, and dynamic time warping, one can match an inaccurate hum to a melody remarkably accurately.
OmniBurst
Burst detection is the activity of finding abnormal aggregates in data streams. Our software, OmniBurst, can detect bursts of varying durations. Our example applications are monitoring gamma rays and stock market price volatility. The software makes use of a shifted wavelet structure to create a linear time filter that can guarantee that no bursts will be missed at the same time that it guarantees (under a reasonable statistical model) that the filter eliminates nearly all false positives.
Title: An Embedded Boundary Integral Solver for the Stokes Equations
Author(s): Biros, George; Ying, Lexing; Zorin, Denis
Abstract:
We present a new method for the solution of the Stokes equations. Our goal is to develop a robust and scalable methodology for two and three dimensional, moving-boundary, flow simulations. Our method is based on Anita Mayo's method for the Poisson's equation: "The Fast Solution of Poisson's and the Biharmonic Equations on Irregular Regions", SIAM J. Num. Anal., 21 (1984), pp. 285--299. We embed the domain in a rectangular domain, for which fast solvers are available, and we impose the boundary conditions as interface (jump) conditions on the velocities and tractions. We use an indirect boundary integral formulation for the homogeneous Stokes equations to compute the jumps. The resulting integral equations are discretized by Nystrom's method. The rectangular domain problem is discretized by finite elements for a velocity-pressure formulation with equal order interpolation bilinear elements (Q1-Q1). Stabilization is used to circumvent the inf-sup condition for the pressure space. For the integral equations, fast matrix vector multiplications are achieved via a N log N algorithm based on a block representation of the discrete integral operator, combined with (kernel independent) singular value decomposition to sparsify low-rank blocks. Our code is built on top of PETSc, an MPI based parallel linear algebra library. The regular grid solver is a Krylov method (Conjugate Residuals) combined with an optimal two-level Schwartz-preconditioner. For the integral equation we use GMRES. We have tested our algorithm on several numerical examples and we have observed optimal convergence rates.
Title: A kernel-independent fast multipole algorithm
Author(s): Biros, George; Ying, Lexing; Zorin, Denis
Abstract:
Title: A kernel-independent fast multipole algorithm (NYU-CS-TR839) Authors: George Biros, Lexing Ying, and Denis Zorin Abstract: We present a new fast multipole method for particle simulations. The main feature of our algorithm is that is kernel independent, in the sense that no analytic expansions are used to represent the far field. Instead we use equivalent densities, which we compute by solving small Dirichlet-type boundary value problems. The translations from the sources to the induced potentials are accelerated by singular value decomposition in 2D and fast Fourier transforms in 3D. We have tested the new method on the single and double layer operators for the Laplacian, the modified Laplacian, the Stokes, the modified Stokes, the Navier, and the modified Navier operators in two and three dimensions. Our numerical results indicate that our method compares very well with the best known implementations of the analytic FMM method for both the Laplacian and modified Laplacian kernels. Its advantage is the (relative) simplicity of the implementation and its immediate extension to more general kernels.
Title: Survey: Eigenvector Analysis in Webpage Rankings
Author(s): Chang, Hung-Hsien
Abstract:
Two major techniques have been proposed for using the structure of links in the World Wide Web to determine the relative significance of Web Pages. The PageRank algorithm \cite{BP98}, which is a critical part of the Google search engine, gives a single measure of importance of each page in the Web. The HITS algorithm \cite{K98} applies to a set of pages believed relevant to a given query, and assigns two values to each page: the degree to which the page is a hub and the degree to which it is an authority. Both algorithms have a natural interpretation in terms of a random walk over the set of pages involved, and in both cases the computation involved amounts to computing an eigenvector over the transition matrix for this random walk.
This paper surveys the literature discussing these two techniques and their variants, and their connection to random walks and eigenvector computation. It also discusses the stability of these techniques under small changes in the Web link structure.
Title: Shrinkage-Based Similarity Metric for Cluster Analysis of Microarray Data
Author(s): Cherepinsky, Vera; Feng, Jiawu; Rejali, Marc; Mishra, Bud
Abstract:
The current standard correlation coefficient used in the analysis of microarray data, including gene expression arrays, was introduced in [1]. Its formulation is rather arbitrary. We give a mathematically rigorous derivation of the correlation coefficient of two gene expression vectors based on James-Stein Shrinkage estimators. We use the background assumptions described in [1], also taking into account the fact that the data can be treated as transformed into normal distributions. While [1] uses zero as an estimator for the expression vector mean μ, we start with the assumption that for each gene, μ is itself a zero-mean normal random variable (with a priori distribution N(0,τ ^{2})), and use Bayesian analysis to update that belief, to obtain a posteriori distribution of μ in terms of the data. The estimator for μ, obtained after shrinkage towards zero, differs from the mean of the data vectors and ultimately leads to a statistically robust estimator for correlation coefficients.
To evaluate the effectiveness of shrinkage, we conducted in silico experiments and also compared similarity metrics on a biological example using the data set from [1]. For the latter, we classified genes involved in the regulation of yeast cell-cycle functions by computing clusters based on various definitions of correlation coefficients, including the one using shrinkage, and contrasting them against clusters based on the activators known in the literature. In addition, we conducted an extensive computational analysis of the data from [1], empirically testing the performance of different values of the shrinkage factor γ and comparing them to the values of γ corresponding to the three metrics adressed here, namely, γ=0 for the Eisen metric, γ=1 for the Pearson correlation coefficient, and γ computed from the data for the Shrinkage metric.
The estimated "false-positives" and "false-negatives" from this study indicate the relative merits of clustering algorithms based on different statistical correlation coefficients as well as the sensitivity of the clustering algorithm to small perturbations in the correlation coefficients. These results indicate that using the shrinkage metric improves the accuracy of the analysis.
All derivation steps are described in detail; all mathematical assertions used in the derivation are proven in the appendix.
[1] Eisen, M.B., Spellman, P.T., Brown, P.O., and Botstein, D. (1998), PNAS USA 95, 14863-14868.
Title: Comparing the Performance of Centralized and Distributed Coordination on Systems with Improved Combining Switches
Author(s): Freudenthal, Eric; Gottlieb, Allan
Abstract:
Memory system congestion due to serialization of hot spot accesses can adversely affect the performance of interprocess coordination algorithms. Hardware and software techniques have been proposed to reduce this congestion and thereby provide superior system performance. The combining networks of Gottlieb et al. automatically parallelize concurrent hot spot memory accesses, improving the performance of algorithms that poll a small number of shared variables. The widely used MCS distributed-spin algorithms take a software approach: they reduce hot spot congestion by polling only variables stored locally. Our investigations detected performance problems in existing designs for combining networks and we propose mechanisms that alleviate them. Simulation studies described herein indicate that a centralized readers writers algorithms executed on the improved combining networks have performance at least competitive to the MCS algorithms.
Title: QTM: Trust Management with Quantified Stochastic Attributes
Author(s): Freudenthal, Eric; Karamcheti, Vijay
Abstract:
Trust management systems enable the construction of access-control infrastructures suitable for protecting sensitive resources from access by unauthorized agents. The state of the art in such systems (i) provide fail-safe in that access will be denied when authorizing credentials are revoked, (ii) can mitigate the risk of insider attacks using mechanisms for threshold authorization in which several independent partially trusted agents are required to co-sponsor sensitive activities, and (iii) are capable of enforcing intra- and inter- organizational access control policies.
Despite these advantages, trust management systems are limited in their ability to express partial trust. Additionally, they are cumbersome to administer when there are a large number of related access rights with differing trust (and thereby access) levels due to the need for explicit enumeration of the exponential number of agent combinations. More importantly, these systems have no provision for fault tolerance in cases where a primary authorization is lost (perhaps due to revocation), but others are available. Such situations may result in a cascading loss of access and possible interruption of service.
In this paper, we propose extending traditional trust management systems through a framework of reliability and confidence metrics. This framework naturally captures partial trust relationships, thereby reducing administrative complexity of access control systems with multiple related trust levels and increasing system availability in the presence of authorization faults while still maintaining equivalent safety properties.
Title: Comparing and Improving Centralized and Distributed Techniques for Coordinating Massively Parallel Shared-Memory Systems
Candidate: Freudenthal, Eric
Advisor(s): Gottlieb, Allan
Abstract:
Two complementary approaches have been proposed to achieve high performance inter-process coordination on highly parallel shared-memory systems. Gottlieb et. al. introduced the technique of combining concurrent memory references, thereby reducing hot spot contention and enabling the bottleneck-free execution of algorithms referencing a small number of shared variables. Mellor- Crummey and Scott introduced an alternative distributed local-spin technique that minimizes hot spot contention by not polling hotspot variables and exploiting the availability of processor-local shared memory. My principal contributions are a comparison of these two approaches, and significant improvements to the former.
The NYU Ultra3 prototype is the only system built that implements memory reference combining. My research utilizes micro-benchmark simulation studies of massively parallel Ultra3 systems executing coordination algorithms. This investigation detects problems in the Ultra3 design that result in higher-than-expected memory latency for reference patterns typical of busy-wait polling. This causes centralized coordination algorithms to perform poorly. Several architectural enhancements are described that significantly reduce the latency of these access patterns, thereby improving the performance of the centralized algorithms.
I investigate existing centralized algorithms for readers-writers and barrier coordination, all of which require fetch-and-add, and discovered variants that require fewer memory accesses (and hence have shorter latency). In addition,my evaluation includes novel algorithms that require only a restricted form of fetch-and-add.
Coordination latency of these algorithms executed on the enhanced combining architecture is compared to the latency of the distributed local-spin alternatives. These comparisons indicate that the distributed local-spin dissemination barrier, which generates no hot spot tra c, has latency slightly inferior to the best centralized algorithms investigated. However, for the less structured readers-writers problem, the centralized algorithms significantly outperform the distributed local-spin algorithm.
Title: Infrastructure Support for Accessing Network Services in Dynamic Network Environments
Candidate: Fu, Xiaodong
Advisor(s): Karamcheti, Vijay
Abstract:
Despite increases in network bandwidth, accessing network services across a wide area network still remains a challenging task. The difficulty mainly comes from the heterogeneous and constantly changing network environment, which usually causes undesirable user experience for network-oblivious applications.
A promising approach to address this is to provide network awareness in communication paths. While several such path-based infrastructures have been proposed, the network awareness provided by them is rather limited. Many challenging problems remain, in particular: (1) how to automatically create effective network paths whose performance is optimized for encountered network conditions, (2) how to dynamically reconfigure such paths when network conditions change, and (3) how to manage and distribute network resources among different paths and between different network regions. Furthermore, there is poor understanding of the benefits of using the path-based approach over other alternatives.
This dissertation describes solutions for these problems, built into a programmable network infrastructure called Composable Adaptive Network Services (CANS). The CANS infrastructure provides applications with network-aware communication paths that are automatically created and dynamically modified. CANS highlights four key mechanisms: (1) a high-level integrated type-based specification of components and network resources; (2) automatic path creation strategies; (3) system support for low-overhead path reconfiguration; and (4) distributed strategies for managing and allocating network resources.
We evaluate these mechanisms using experiments with typical applications running in the CANS infrastructure, and extensive simulation on a large scale network topology to compare with other alternatives. Experimental results validate the effectiveness of our approach, verifying that (1) the path-based approach provides the best and the most robust performance under a wide range of network configurations as compared to end-point or proxy-based alternatives; (2) automatic generation of network-aware paths is feasible and provides considerable performance advantages, requiring only minimal input from applications; (3) path reconfiguration strategies ensure continuous adaptation and provide desirable adaptation behaviors by using automatically generated paths; (4) both run-time overhead and reconfiguration time of CANS paths are negligible for most applications; (5) the resource management and allocation strategies allow effective setting up shared resource pools in the network and sharing resources among paths.
Title: Why Path-Based Adaptation? Performance Implications of Different Adaptation Mechanisms for Network Content Delivery
Author(s): Fu, Xiaodong; Karamcheti, Vijay
Abstract:
Because of heterogeneous and dynamic changing network environments, content delivery across the network requires system support for coping with different network conditions in order to provide satisfactory user experiences. Despite the existence of many adaptation frameworks, the question that which adaptation approach performs the best under what network configurations still remains unanswered. The performance implication of different adaptation approaches (end-point, proxy-based and path-based approaches) has not been studied yet. This paper aims to address this shortcoming by conducting a series simulation-based experiments to compare performance among these adaptation approaches under different network configurations. In order to make a fair comparison, in this paper approach-neutral strategies are proposed for constructing communication paths and managing network resources. The experiment results show that there are well-defined network environments under which each of these approaches delivers its best performance; and among them, the path-based approach, which uses the whole communication path to do adaptation, provides the best and the most robust performance under different network configurations, and for different types of servers and clients.
Title: Balancing Neumann-Neumann Preconditioners for the Mixed Formulation of Almost-Incompressible Linear Elasticity
Author(s): Goldfeld, Paulo
Abstract:
Balancing Neumann-Neumann methods are extended to the equations arising from the mixed formulation of almost-incompressible linear elasticity problems discretized with discontinuous-pressure finite elements. This family of domain decomposition algorithms has previously been shown to be effective for large finite element approximations of positive definite elliptic problems. Our methods are proved to be scalable and to depend weakly on the size of the local problems. Our work is an extension of previous work by Pavarino and Widlund on BNN methods for Stokes equation.
Our iterative substructuring methods are based on the partition of the unknowns into interior ones - including interior displacements and pressures with zero average on every subdomain - and interface ones - displacements on the geometric interface and constant-by-subdomain pressures. The restriction of the problem to the interior degrees of freedom is then a collection of decoupled local problems that are well-posed even in the incompressible limit. The interior variables are eliminated and a hybrid preconditioner of BNN type is designed for the Schur complement problem. The iterates are restricted to a benign subspace, on which the preconditioned operator is positive definite, allowing for the use of conjugate gradient methods.
A complete convergence analysis of the method is presented for the constant coefficient case. The algorithm is extended to handle discontinuous coefficients, but a full analysis is not provided. Extensions of the algorithm and of the analysis are also presented for problems combining pure-displacement and mixed finite elements in different subregions. An algorithm is also proposed for problems with continuous discrete pressure spaces.
All the algorithms discussed have been implemented in parallel codes that have been successfully tested on large sample problems on large parallel computers; results are presented and discussed. Implementations issues are also discussed, including a version of our main algorithm that does not require the solution of any auxiliary saddle-point problem since all subproblems of the preconditioner can be reduced to solving symmetric positive definite linear systems.